perf(p2p-14): add adaptive read-path bypasses (#671)

[jlon]

Summary

• Add adaptive read-side optimizations after the data-plane fast path is isolated.
• Fold the reduced-intrusion follow-up that extracts the internal read-acceleration boundary, while keeping filesystem API narrowing and harness migration out of this PR.

Design

• Keep adaptive behavior and read-path internals together: the same slice now owns adaptive EWMA state, read-flight coordination, and local accelerated-cache state.
• Preserve the existing read order and worker fallback semantics; telemetry expansion still lands later in p2p-16.

Implementation

• Add curvine-client/src/file/read_accelerator.rs and move read-chunk cache, read-flight locks, adaptive bypass state, and P2P read attempts behind that boundary.
• Rewire block_reader, fs_context, and fs_reader_base to use the new boundary without changing the worker fallback path.
• Keep only the minimal compatibility hooks needed so later slices can narrow the public filesystem surface cleanly.

Verification

• cargo clippy -p curvine-client --all-targets -- -D warnings
• cargo test -p curvine-client --lib read_accelerator -- --nocapture
• cargo test -p curvine-client --lib block_reader -- --nocapture

Review Note

• This split series still targets main, so GitHub shows cumulative diff.
• The intended incremental scope for this PR is the delta from p2p-13 to p2p-14: adaptive read-path work plus the internal read-acceleration boundary extraction.

[jlon]

Part 14/17 of the split P2P series that supersedes #671.

This PR targets , so the visible diff is cumulative. Review the slice described in this PR body instead of treating the whole diff as new.

Intended review delta is p2p-14 on top of #750.

[jlon]

Review note: this is part 14/17 of the split P2P series that supersedes #671.

This PR targets main, so the visible diff is cumulative. Please review the slice described in this PR body.

Review after #750.