perf: optimize log list/export with 3-layer metadata loading

[ntwrknrd]

Problem

nssh log list and nssh log export freeze for 30+ seconds when there are many recording files or large recordings.

Root Cause

ReadCastMetadata() was reading the entire cast file to find the final timestamp for duration calculation. With large recordings (multi-hour sessions), this meant reading hundreds of MB per file.

Solution

Implemented a 3-layer optimization strategy:

Layer 1: Index-first lookup

• Check .index.json sidecar file first (already exists, ~1KB)
• Contains StartedAt/FinishedAt timestamps - no file parsing needed

Layer 2: Tail-read fallback

• For files without index, read last 64KB instead of entire file
• Parse backwards to find final timestamp
• Reduces O(file_size) → O(1)

Layer 3: Lazy loading with mtime pre-sort

• New ListCastFilesWithMtime() collects paths + mtimes cheaply
• Sort by mtime before loading any metadata
• IterSessionRecordsLimit() only parses top N files when --last is used

Performance Impact

Scenario	Before	After
nssh log list --last 5	Loads all sessions	Loads only 5
Files with .index.json	Full file scan	~1KB JSON read
Files without index	Full file scan	64KB tail read

Testing

• make build ✓
• make test ✓
• go vet ./... ✓