[ecbadeaux]

Replace regex tag sanitizer with a table-driven one

Summary

MeterId built its spectatord id by running std::regex_replace on the name and on every tag key and value, then joining them with an ostringstream. This was the dominant cost of r.CreateCounter(name, tags). This PR replaces it with a compile-time lookup table and a single pre-sized string.

• AppendSanitized writes each byte as kValidChars[byte] ? byte : '_'. The 256-entry table is built by a constexpr function, so nothing runs at startup. The old global std::regex is gone.
• ToSpectatorId computes the exact output size, reserves once, and appends with no ostringstream.
• Helpers are in an anonymous namespace, so there is no new public API and no version bump is needed. meter_id.h still includes <regex> so code that relied on it keeps compiling.
• Emitted lines are byte-for-byte unchanged.

Why this is faster than the per-character version

Google Benchmark on an id with a 35-character name and 7 tags:

Variant	Time
Original (regex + ostringstream)	10,905 ns
Pre-sized string, per-character += with range checks	370 ns
This PR: resize once + lookup table	130 ns

Both new versions reserve the output up front, so neither reallocates. The difference is what happens per character.

• Per-character +=: reserving does not change the size, so every append still increases the string's size by one. Each iteration writes the new size back, stores the byte, rewrites the trailing NUL, and re-checks short-string vs heap mode. The generated code makes an out-of-line push_back call per character for the name.
• This PR: the string is resized to its final length before the loop, since sanitizing never changes length. The loop only reads a byte, looks it up, and writes a byte through a raw pointer. It never touches the size field, and the compiler unrolls it.
• Range checks vs table: the table replaces about eight comparisons per byte with one load.