Performance tweaks

[j-rafique]

• 1. supernode/cascade/download.go: Caps how many candidate keys get passed into P2PClient.BatchRetrieveStream (keyCap = max(need*50, 5000)) and stops building remainingKeys once that cap is reached; rationale is to avoid DHT per-key preprocessing cost on huge key lists when only a small top-up is needed; impact is noticeably lower CPU/alloc/GC late in the download when need is small (possible extra attempt if many keys are missing/unavailable).
• 2. supernode/cascade/download.go: Raises the initial decode target from 17% to 20% (targetRequiredPercent) and skips RQ.DecodeFromPrepared until have >= reqCount; rationale is to avoid expensive decode attempts at borderline/under-target coverage; impact is fewer wasted decode cycles and typically faster time-to-success, with a small increase in symbols fetched up front.
• 3. supernode/cascade/download.go: Uses a prepared workspace + streaming retrieval with incremental retries (attempt loop with reqCount increasing by ~5%); rationale is to reduce memory usage and improve resilience vs single-shot decode; impact is higher success rate and often better overall download time, though worst-case can do multiple retrieve+decode rounds (more CPU/network than fail-fast).
• 4. supernode/adaptors/p2p.go: Derives symbol keys from codec.Layout (symbolKeysFromLayout) instead of recursively walking the symbols directory; rationale is to avoid expensive filesystem traversal; impact is materially less disk I/O and faster first-pass store (risk only if layout and on-disk symbols diverge).

[roomote]

Rooviewer   See task on Roo Cloud

Review complete. No issues found.

This PR continues the performance improvements with two additional commits:

• Allocation reduction in closest-node selection: Pre-computes ignored node set once for batch operations, pre-allocates slices based on total node count, and releases locks earlier (before sorting) to reduce contention
• Synchronous StoreBatch for original data: Ensures original/local batches are durably committed before returning, preventing data loss if callers delete source files immediately after StoreBatch returns

Previous reviews

• 74e0dcf: Review #1

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}