perf: fast paths for Thread.current, Fiber.current, Mutex#synchronize

[linyiru]

Part of #378.

shape	rubyrs	CRuby 3.4	×
Thread.current	150	25	6
Thread.current[:k]	248	27	9
Mutex#synchronize {}	1257	53	24

Per Rails request there are about 20 Thread.current and about 15 Fiber.current calls. IsolatedExecutionState, CurrentAttributes, the logger and Concurrent::Map all go through them. Monitor/Mutex sit on the logger and cache paths.

Do:

• Profile each path with sample.
• Serve Thread.current and Fiber.current as cached handles (no lookup or allocation per call).
• Make Thread#[] / #[]= direct fiber-local table reads.
• Make Mutex#synchronize a native lock/yield/ensure-unlock, with no Ruby-level begin/ensure frame.

Done when:

• Each shape is within 2× of CRuby.
• Existing Thread/Fiber/Mutex specs pass.
• A diff_cruby fixture covers fiber-local vs thread-local semantics.

[linyiru]

Reopening. This issue was closed when #395 merged, but #395 said it should stay open, and the "each shape within 2× of CRuby" condition is still not met.

#390 has since landed and replaced #395's implementation (see the note at the top of #390). These are its numbers from the PR, ns/op on the Mac Studio M2 Max, interpreter:

shape	before	after #390	CRuby 3.4.8	× CRuby
Thread.current	107.5	27.9	16.6	1.7 ✅
Thread.current[:k] (shape-bench)	242.0	67.8	19.7	3.4
Thread.current[:k]=	288	99	32	3.1
Fiber.current	252	41	17	2.4
Mutex lock+unlock	779	96	36	2.7
Mutex#synchronize {} (shape-bench)	1225.6	444.1	53.8	8.3

Status of the done-when items:

• Within 2× of CRuby: met only for Thread.current. The rest now sit on the interpreter's generic call floor (constant load, builtin call, Hash lookup, block invoke), not in these methods.
• Existing Thread/Fiber/Mutex specs pass: yes.
• A diff_cruby fixture covers fiber-local vs thread-local semantics: yes (thread_fiber_locals.rb).

The remaining gap is tracked by:

• perf(hash): small-table linear lookup for Symbol/immediate keys #388: small-table Hash lookup, for Thread.current[:k].
• perf(jit): compile Thread.current / Thread#[] / Mutex intrinsics inline #389: JIT-inline these intrinsics. The measured ideal is below CRuby.
• perf: interpreter block invocation from native iterators #382: block invocation, the floor under synchronize.

Close this issue once those bring the remaining shapes within 2×.