feat: add byte-level progress reporting with EWMA throughput tracking

[deepjoy]

Summary

• Add opt-in byte-level progress reporting for streaming transfers (uploads, downloads, file copies) with lock-free atomic counters on the executor hot path
• Introduce a background progress ticker that emits TaskProgress events on a dedicated broadcast channel with EWMA-smoothed throughput (bytes/sec) and estimated time remaining
• Expose TaskContext::set_bytes_total / add_bytes / report_bytes convenience API for executors, and Scheduler::subscribe_progress / byte_progress for consumers

Details

Core mechanism: Each active task gets progress_bytes and total_bytes AtomicU64 fields on its IoTracker. Executors update these via TaskContext::add_bytes() — no locks, no allocations, no broadcast on the hot path.

Progress ticker: When SchedulerBuilder::progress_interval(duration) is set, a background task polls byte counters at the configured interval, computes EWMA-smoothed throughput (alpha=0.3), and emits TaskProgress events on a dedicated channel (Scheduler::subscribe_progress()). Stale trackers are pruned automatically when tasks complete.

Zero cost when disabled: The ticker is opt-in — when progress_interval is not set, the only cost is two unused AtomicU64 fields per active task. Benchmarks confirm no measurable overhead.

Files changed

Area	Files
Atomics	registry/io_tracker.rs — progress_bytes, total_bytes fields + snapshot API
Executor API	registry/context.rs — set_bytes_total, add_bytes, report_bytes on TaskContext
Reporter	scheduler/progress.rs — ProgressReporter byte methods, ThroughputTracker (EWMA), run_progress_ticker, TaskProgress struct
Active tracking	scheduler/dispatch.rs — ByteProgressSnapshot type alias, byte_progress_snapshots()
Scheduler wiring	scheduler/mod.rs, run_loop.rs, builder.rs, event.rs, queries.rs
Tests	scheduler/tests.rs — 3 new integration tests
Benchmarks	benches/scheduler.rs — byte_progress_overhead, byte_progress_snapshot
Docs	lib.rs, scheduler/mod.rs, scheduler/progress.rs — crate and module-level docs

Benchmarks confirm zero overhead (noop vs byte-reporting within noise)

Fixes #12