Add regex/grep support to HypGrep

[platypii]

Make HypGrep accept a RegExp as the query and surface a grep-style CLI, while keeping per-query transfer bytes bounded via index pruning.

Key changes

• New src/regex.js: extracts mandatory literal substrings from a regex source in DNF form (handles escapes, character classes, groups, quantifiers, anchors, top-level alternation).
• queryIndex accepts RegExp: each top-level alternation branch becomes its own n-gram conjunction; a block is a candidate if any branch is fully satisfied (Zoekt-style branch ORing).
• parquetFind / parquetSearch accept RegExp directly — default row predicate becomes regex.test(value); scorer for parquetSearch uses the regex's own flags.
• Full-scan fallback when a query yields no extractable n-grams (short string, /./, /foo|bar/ where a branch matches anything) — correct grep semantics, no index acceleration.
• New hypgrep search CLI subcommand: npx hypgrep search file.parquet '<pattern>', with --limit N, --index <path>, -c/--count, -i/--ignore-case. Patterns wrapped as /regex/flags are parsed as regex; exit code 1 on no matches (grep convention).
• Fixed an unsafe over-extraction across top-level alternation (/foo|bar/ previously yielded ['foo']).
• ARCH.md + README updated with regex examples, CLI usage, and the new limitation list.

Breaking changes

• extractRegexLiterals now returns string[][] (DNF) instead of string[]. Internal only — not part of the public API.

Known limitations (documented in ARCH.md)

• Alternation inside groups under-extracted (/abc(foo|bar)def/ yields ['abc','def'], not the tighter DNF). Group recursion is the next step.
• Per-row precision still missing — candidate blocks are scanned in full.
• DNF size has no cap yet; deeply nested alternation could blow up once group recursion lands.