Fix inverted offset check in encode_into - #1310

Encoder.encode_into(msg, buf, offset) resizes buf down to max(8, 1.5 * offset) whenever offset < len(buf), so passing an offset throws away the capacity of a preallocated buffer. Leaving the offset out keeps it:

import msgspec

enc = msgspec.msgpack.Encoder()
buf = bytearray(1000)                      # __sizeof__() == 1057
enc.encode_into({"key": "value"}, buf, 4)
buf.__sizeof__()                           # 79 (still 1057 without the offset)

The length-prefix example in the perf tips uses this pattern (bytearray(64) then encode_into(msg, buffer, 4)), and the docstring says the underlying buffer won't be truncated.

When offset > len(buf) nothing is resized up front. The buffer is grown later by ms_resize_bytearray, and the bytes between the old end and offset are left uninitialized (encode_into(1, bytearray(b"abc"), 64) leaves non-zero leftover bytes in buf[3:64]).

#753 meant to grow the buffer when the offset is past the end, but the check reads offset < buf_size. This flips it to > and zero-fills the gap, so in-bounds offsets leave the buffer alone.

Tests: in test_encode_into_offset (json and msgpack) the out-of-bounds case now checks the gap is zeros, and a new case checks that an in-bounds offset keeps the capacity. Both fail on main. pytest tests/unit passes on Python 3.13 (6582 passed, 158 skipped), and ruff check, ruff format --diff and codespell are clean.

#1160 changes the same function but keeps this block as it is, so whichever goes in second gets a small conflict.