Fixes #339

### Motivation

Currently the relative path is used in unit tests to specify the token
file (`.test-token.txt`) and the credential directory (`./test-conf`).
If we don't run the `pulsar-tests` binary in a subdirectory, it won't
be able to read these files.

### Modifications

Add the macro `-DTOKEN_PATH="..." -DTEST_CONF_DIR="..."` to specify
absolute paths according to the `PROJECT_SOURCE_DIR`, which is defined
by CMake as the absolute path of the project directory.

### Verifications

```bash
cmake -B build
cmake --build build -j8
# Run the test in project directory
./build/tests/pulsar-tests --gtest_filter='AuthPlugin*:*testRSAEncryption:*testEncryptionFailure:*DecryptionFailedMessages'
# Run the test in a subdirectory
cd build
./tests/pulsar-tests --gtest_filter='AuthPlugin*:*testRSAEncryption:*testEncryptionFailure:*DecryptionFailedMessages'
```

(cherry picked from commit 47fb809)

)

Fixes #88

### Motivation

When `close` is called if the consumer has already called `unsubscribe`
or `close`, it should not fail. See

https://github.com/apache/pulsar/blob/428c18c8d0c3d135189920740192982e11ffb2bf/pulsar-client/src/main/java/org/apache/pulsar/client/impl/ConsumerImpl.java#L1034

### Modifications

Use the same close logic with Java client.

Add `testCloseAgainBeforeCloseDone` and `testCloseAfterUnsubscribe` to
verify the new behaviors of `Consumer::close`.

(cherry picked from commit a8402da)

Fixes #341

### Motivation

When a lazy partitioned producer sends two messages, the flow is:
1. `start` is called to grab the connection via `grab()`.
2. Generate 0 as the sequence id of the 1st message.
3. Add the 1st message into the queuea.
4. The connection is established, `msgSequenceGenerator_` is reset from
   1 to 0.
5. When sending the 2nd message, 0 is also generated as the sequence id.

Then two messages have the same sequence id.

### Modifications

For lazy partitioned producers, if the internal producer is not started,
sending the message in the callback of its future.

Add `ChunkDedupTest#testLazyPartitionedProducer` to verify it since only
the `tests/chunkdedup/docker-compose.yml` enables the deduplication.

(cherry picked from commit bb16f24)

Fixes #346

### Motivation

#336 changes the key of
the `ClientConnection` in `ConnectionPool`, while in
`ClientConnection::close`, it still passes the old key (logical address)
to `ConnectionPool::remove`, which results in the connection could never
be removed and destroyed until being deleted as a stale connection.

What's worse, if the key does not exist, the iterator returned by
`std::map::find` will still be dereferenced, which might cause crash in
some platforms. See
https://github.com/apache/pulsar-client-cpp/blob/8d32fd254e294d1fabba73aed70115a434b341ef/lib/ConnectionPool.cc#L122-L123

### Modifications

- Avoid dereferencing the iterator if it's invalid in
  `ConnectionPool::remove`.
- Store the key suffix in `ClientConnection` and pass the correct key to
  `ConnectionPool::remove` in `ClientConnection::close`
- Add `ClientTest.testConnectionClose` to verify
  `ClientConnection::close` can remove itself from the pool and the
  connection will be destroyed eventually.

(cherry picked from commit 6d47e94)

(cherry picked from commit c771b12)

)

Fixes #348

Fixes #349

### Motivation

When `close` is called in `ClientConnection`'s constructor,
`shared_from_this()` will be called, which results in a
`std::bad_weak_ptr` error. This error does not happen before
#317 because
`shared_from_this()` could only be called when the `producers` or
`consumers` field is not empty.

### Modifications

Throw `ResultAuthenticationError` when there is anything wrong with
authentication in `ClientConnection`'s constructor. Then catch the
result in `ConnectionPool::getConnectionAsync` and returned the failed
future.

In addition, check `authentication_` even for non-TLS URLs. Otherwise,
the null authentication will be used to construct `CommandConnect`.

Add `testInvalidPlugin` and `testTlsConfigError` to verify the changes.

(cherry picked from commit 7bb94f4)

Fixes #352

### Motivation

#303 adds the flush
callback to the last `OpSendMsg` instead of adding to the batch message
container. However, `batchMessageAndSend` will create an `OpSendMsg` and
add it to the `pendingMessagesQueue_`.

https://github.com/apache/pulsar-client-cpp/blob/7bb94f45b917ed30b5302ac93ffa1f1942fc6313/lib/ProducerImpl.cc#L384-L389

In the code above, `pendingMessagesQueue_` could never be empty and the
callback will be added again by `opSendMsg->addTrackerCallback`. The 1st
time it's added in `createOpSendMsg` or `createOpSendMsgs` called by
`batchMessageAndSend`.

### Motivation

Add the callback to the last `OpSendMsg only when the batch message
container is empty.

In `testFlushBatch`, replace the `flush` call with the `flushAsync` call
and verify the callback is only called once after it's completed.

(cherry picked from commit 37ea769)

…n macOS (#354)

### Motivation

#290 brings a regression
that on macOS, Protobuf is always found with CMake Config mode, which
does not set the `Protobuf_LIBRARIES` variable so that the
libpulsarwithdeps.a misses the symbols of Protobuf.

### Modifications

When `LINK_STATIC` is ON, use CMake Module mode to find the Protobuf.

Add `build-static-library.sh` to build libraries with static
dependencies and verify these libraries in PR workflow. Upload the
pre-built binaries in the build workflow.

(cherry picked from commit f75b39b)