Follow the ensure codegen of mruby-compiler, and fix a pending return past the frame - #313

This PR contains three patches.

Two of them catch up with a change in upstream mruby-compiler; the other fixes a long-standing bug in mruby/c itself.
They are in one PR because neither passes CI on its own.

Let OP_RESCUE answer false for a non-exception value

Follows mruby/mruby@b7df80f ("mruby-compiler: name the exception as $! in a rescue modifier and an ensure"), which picoruby picked up through its mruby-compiler sync.

The codegen now emits OP_RESCUE at the entry of an ensure clause to decide whether $! should name the exception it is unwinding:

EXCEPT   R3            ; nil on a normal entry
GETGV    R4  $!
OCLASS   R5
GETMCNST R5  ::Exception
RESCUE   R3  R5        ; R3.is_a?(Exception)?
JMPNOT   R5  ...
SETGV    $!  R3

So R[a] of OP_RESCUE is no longer always an exception: it is nil when the clause is entered normally, or a break/return passing through.
The mruby VM was changed accordingly (!mrb_break_p(exc) && ...), while op_rescue in mruby/c asserted MRBC_TT_EXCEPTION and aborted every begin ... ensure ... end on the normal path:

femtoruby: vm.c:1039: op_rescue: Assertion `mrbc_type(regs[a]) == MRBC_TT_EXCEPTION' failed.

This is backward compatible.
An mrbc from before that change (at mruby 4.0) emits OP_RESCUE only inside a rescue clause, where R[a] is always an exception, so the new check is always true there and the result is unchanged.

Carry a pending return's value in a box, not past the frame

This is an existing bug, independent of the compiler: bytecode from an mrbc (even in mruby 4.0) crashes the same way.

When a return has an ensure clause to run, sub_op_return parked the return value in regs[nregs], the register just past the current frame, and jumped to the clause.
But a method called from the ensure clause builds its frame exactly there, so anything but a C function in the clause clobbered the value:

def m
  return 42
ensure
  puts "bye"   # Kernel#puts is Ruby: its frame overlaps regs[nregs]
end
p m
# => class.h:277: mrbc_find_class_by_object: Assertion `!"Invalid value type."' failed.

This patch solves that by introducing release_pending_return() (see diff)

test/ensure_return_test.rb covers a method call in the ensure clause, a local rewritten by the clause, two ensure clauses, a nested pending return, a raise replacing the return, an object value, and $! (now it is set by mruby-compiler) in an ensure clause entered normally, by an exception, and by a break.

Add Object#__pat_values for a hash pattern's type check

Follows mruby/mruby@b2d2cd1 ("mruby-compiler: hold #deconstruct_keys to the Hash it has to answer"). The compiler now sends __pat_values to whatever #deconstruct_keys answered, even for a pattern with no keys, and relies on the receiver's class for the type check: Hash#__pat_values answers the values, Object#__pat_values raises TypeError: deconstruct_keys must return Hash.

Backward compatible as well: an older mrbc never sends __pat_values to anything but a Hash.