Optimize slog dispatch for spawned cache tasks

[proboscis]

Summary

• add a WriterHandler fast-path for PyTell dispatch so cache slog() calls can write directly into the Rust store without re-entering the full handler path
• cache handler trace metadata on prompt segments and reuse it during dispatch/trace assembly instead of recomputing debug_info() on every scan
• dedupe the shared visible handler prefix when resuming unstarted continuations created from GetHandlers()/Spawn, which removes the duplicated handler chain seen in cached spawned tasks
• add a regression test that compares the handler stack seen by a cached direct call vs the same cached call through Spawn

Verification

• uvx maturin develop --release
• uv run pytest tests/test_cache_await_spawn_hang.py -q
  • 7 passed in 5.52s
• uv run pytest tests/test_cache_await_spawn_hang.py tests/test_with_handler_type_filter_regression.py tests/effects/test_local_handler_ask.py::test_local_eval_keeps_typed_handler_filtering -q
  • 20 passed in 6.65s
• cache+slog benchmark (N=500, persistent sqlite cache, same cache task with cache slog monkeypatched on/off)
  • cache_slog_off ok= True elapsed= 1.318 len= 500
  • cache_slog_on ok= True elapsed= 2.089 len= 500
• spawned cached-task handler probe
  • GetHandlers() length inside the spawned cached task is now 9 instead of the duplicated 18
  • direct cached call and spawned cached call now report the same handler stack

Notes

• cargo test in packages/doeff-vm-core passed.
• cargo test in packages/doeff-core-effects is currently blocked by pre-existing scheduler test compile errors that still reference ReadySet::Fifo/ReadySet::Priority; this is unrelated to the writer-dispatch fix.

Issue: slog-handler-perf-concurrent-spawn

[proboscis]

Addressed review round 1:

• replaced WriterHandler string matching with a typed builtin discriminant (BuiltinHandlerKind::Writer)
• replaced PyTell name matching with an isinstance(doeff_vm.PyTell) check backed by a cached Python type object
• guarded unstarted-continuation dedupe with if !handlers.is_empty() and documented that prefix trimming is strictly position-sensitive
• added an explicit comment that fast-path type-filter misses must fall back to full dispatch semantics
• made the fast-path effect object conversion lazy so the common no-type-filter path avoids the extra Python conversion

Re-verified after the follow-up fix:

• uvx maturin develop --release
• uv run pytest tests/test_cache_await_spawn_hang.py tests/test_with_handler_type_filter_regression.py tests/effects/test_local_handler_ask.py::test_local_eval_keeps_typed_handler_filtering -q -> 21 passed in 7.36s
• cargo test in packages/doeff-vm-core -> pass
• benchmark: cache_slog_off 1.41s, cache_slog_on 2.162s for N=500

[proboscis]

Round 3 addressed in 0ac9ceb.

Changes:

• removed the writer fast-path entirely (including the builtin writer discriminant plumbing)
• removed the Arc::ptr_eq shared-prefix dedup
• kept GetHandlers() full-stack semantics, but changed spawned continuation capture to store an inherited handler-prefix descriptor instead of rebasing resume onto a different segment
• for SpawnAwaitHandlers, CreateContinuation now records the inherited visible handler prefix and an offset so prepended local handlers like the blocking sync_await_handler remain local instead of being skipped

Why this fixes the root cause:

• the continuation still resumes relative to the actual resume site, so return/delegate topology stays intact
• only the inherited portion of the captured handler list is elided at resume, and only after the local-prefix offset, so we no longer duplicate the caller-visible chain or accidentally drop task-local handlers

Validation:

• uvx maturin develop --release
• uv run pytest tests/test_cache_await_spawn_hang.py -q -> 7 passed
• uv run pytest tests/test_cache_await_spawn_hang.py tests/test_with_handler_type_filter_regression.py tests/effects/test_local_handler_ask.py::test_local_eval_keeps_typed_handler_filtering -q -> 21 passed
• cargo test in packages/doeff-vm-core -> pass
• benchmark rerun after removing the writer fast-path (N=500):
  • cache_slog_off ok=True elapsed=2.400 len=500
  • cache_slog_on ok=True elapsed=2.456 len=500