perf: split Publisher::send() into separate prefix/value parts

[akonradi]

Avoids a redundant copy: the previous code assembled a complete 'prefix+value' line in a thread-local scratch std::string, then copied that whole line into the Unix-domain transport's accumulation buffer -- the prefix got copied twice per measurement. Publisher::send() now takes the cached per-meter prefix and the freshly-formatted value as two separate string_view parts; the Unix-domain sender appends each straight into its own buffer_, so the prefix is copied into its final resting place exactly once. StatelessMeter::send()/send_uint() format the value directly into a plain stack array (no thread-local std::string staging buffer needed at all now, since there's no longer a combined line to assemble) and pass it on as a view.

UDP sends the two parts as one datagram via a 2-buffer asio scatter/gather write -- no scaling concern there, since each UDP send is always exactly these two buffers, never batched with others the way Unix-domain measurements are.

No change to the public Id/Tags/Counter/Gauge/.../Registry API.

Microbenchmark (300 Counters, 500K Increment() calls, real Unix-domain receiver): ~0.050us/call vs baseline's ~0.056us/call, ~10% faster

[akonradi]

cc @jasonk000

[jasonk000]

@akonradi looks like a rebase should fix the build issue, can you try that? thx

[jasonk000]

A couple of suggestions.