perf: speed up session scanning (minimal change)

[ScriptedAlchemy]

Goal: improve performance without changing outputs or data shapes.

Changes (intentionally minimal / compatible):

• getCodexFirstPromptTimestamp now streams ~/.codex/history.jsonl instead of readFile().
• Adds a fast pre-filter to skip JSON.parse for irrelevant lines.
• Adds bounded concurrency across session files (min(cpuCount, 8)).
• Preserves existing CodexUsageData output (including events + sorting).

Perf on this machine (2025 sessions ~1.1GB, 3706 files):

• collectCodexUsageData(2025) ~3.0s (was ~6.7s).

[ScriptedAlchemy]

Reviewed.

• Scope looks minimal: only src/collector.ts changes; CodexUsageData shape + events sorting preserved; stats.ts untouched.
• Perf wins come from (1) streaming history.jsonl instead of readFile(), (2) skipping JSON.parse for irrelevant lines, and (3) bounded file-level concurrency.

Notes:

• Concurrency is the only meaningful behavioral change (IO pattern). min(cpuCount, 8) seems reasonable; if we want to be extra conservative for slower disks, we could cap at 4.
• getTopLevelType + the "payload" guard is a good safety to avoid matching nested "type" fields.

Local perf on this machine (2025 sessions ~1.1GB, 3706 files): collectCodexUsageData(2025) ~3.0s (was ~6.7s).