skills: generalize --tail to numeric line counts with k/w shorthand and remove time-window tails

[mouxinqq]

Motivation

• Unify --tail semantics across skills so it only represents a line-count and avoid ambiguous minute/hour/day tail syntax.
• Make shorthand parsing generic (e.g. 1k/1w) instead of special-casing particular values like 2k.
• Keep time-based filtering explicit via --start/--end to avoid implicit time-window tail behavior.

Description

• Reworked --tail parsing in stat_cache_hitrate.py to use a single regex ^(\d+)([kw])?$ and convert k→1000 and w→10000, returning an integer line count and removing minute/hour/day parsing paths.
• Applied the same generic --tail parsing in troubleshoot.py, added the missing import re, and made the parser return { "type": "lines", "value": <int> } for downstream use.
• Removed/streamlined time-window branches: deleted minute-based tail filtering code and adjusted read_lines, grep_and_parse, and grep_count to always treat --tail as a line count when present.
• Updated CLI help text and both SKILL.md files to document that --tail is line-count only (examples include 5000, 1k, 1w) and that time filtering must use --start/--end.

Testing

• Ran python3 -m py_compile on both scripts (stat_cache_hitrate.py and troubleshoot.py) and compilation succeeded.
• Executed module-level smoke checks that call the new parse_tail_arg with inputs 2000, 1k, and 1w for both scripts, and they returned [2000, 1000, 10000] (tests passed).
• An initial import smoke test failed due to cross-module import ordering, which was diagnosed and resolved before the final successful tests.

---

Codex Task