perf(runtime): read schema graph and subschema caches without locking

[tanmaykm]

Closes #128.

• _schema_at took spec.graph_lock on every lookup to read the plain Dict spec.subschemas, which serialized concurrent decoders.
• There was also a data race: _schema_graph called haskey(spec.graphs, ...) outside the lock while another thread could be writing to it under the lock.

Both caches now live in a SchemaCache with @atomic fields that hold immutable Dict snapshots:

• Cache hits are lock-free atomic loads.
• A miss takes graph_lock, re-reads the current snapshot, and returns any entry a racing writer has already published. Otherwise it copies the snapshot, inserts, and publishes the copy. A published Dict is never mutated.
• Graphs are still compiled once per direction (checked again under the lock).
• The Spec keyword constructor that generated modules call is unchanged, so no contract bump is needed.

Rough numbers on Julia 1.12 with 4 threads:

• An uncontended _schema_at hit went from 36 ns to 24 ns.
• With 4 tasks contending, the amortized cost per call went from 118–146 ns to 9–11 ns.

Tests

• Tests for cache identity across directions, and a cold-cache multi-threaded stress test over the runtime and a generated client's decoding.
• CI change: JULIA_NUM_THREADS: 4 is now set for the test job, so the stress test runs threaded (Pkg.test forwards the thread count). This applies to every test file. AGENTS.md documents it.
• Full suites pass on Julia 1.12 and 1.10 with 4 threads.

Possible follow-ups (minor review notes)

• Add a test showing that a cached hit completes while graph_lock is held by another task. That would catch a future change that adds the lock back to the hit path.
• Use acquire loads and release stores instead of the default seq_cst ones.