[Feature] Add per-task and partial-scope granularity to scope_stats collection

[ChaoZheng109]

Follow-up to #858 / #573.

Summary

scope_stats (#858) profiles ring-buffer (and tensormap) resource peaks at per-scope granularity. This issue tracks extending it with per-task and partial-scope collection granularity, added incrementally so the hot path stays cheap.

Motivation / Use Case

#858 deliberately samples only at scope_begin / scope_end to keep overhead minimal — one bool load when disabled, two pure-value calls per scope when enabled. That answers "which scope drove the peak" but not:

• Per-task attribution — within a scope that dispatches many tasks, which task(s) actually pushed ring task-window / heap to the high-water mark.
• Partial-scope sampling — intermediate peaks inside a long-running scope, so a single scope record does not flatten a time-varying occupancy profile into one number.

Without these, debugging memory pressure inside a large scope still requires guesswork once scope_stats has narrowed it to a scope.

Proposed API / Behavior

• Add an opt-in finer sampling level on top of --enable-scope-stats (e.g. a per-task probe at submit_task and/or periodic in-scope sampling), so the default per-scope mode keeps its current low overhead and the finer mode is paid for only when explicitly enabled.
• Emit the extra granularity into the existing scope_stats.jsonl stream (additional record types / fields) rather than a new artifact, to keep the consumer side unified.
• Preserve the single-producer, non-atomic, pooled-buffer + host-drain model from Add: scope_stats collector for per-scope queue-fill peaks #858 so the finer sampling does not add atomics or fixed-ring back-pressure on the hot path.

Alternatives Considered

• Always-on per-task sampling — rejected: adds per-submit_task cost to every profiled run; Add: scope_stats collector for per-scope queue-fill peaks #858 intentionally avoided this.
• A separate per-task profiler unrelated to scope_stats — rejected: duplicates the buffer-pool/drain infra and the JSON consumer.

Additional Context

• Implemented baseline: Add: scope_stats collector for per-scope queue-fill peaks #858 (docs/dfx/scope-stats.md), merged as 22538de.
• Originating request: [Feature] Add ring buffer memory usage profiling #573 (closed; per-scope requirements satisfied, per-task deferred here).