Keep chars_format::javascript out of the common hot path

[Algunenano]

chars_format::javascript, new in 8.3.0, adds its ECMAScript leading-zero rule as a runtime test on fmt inside parse_number_string. Unlike basic_json_fmt, which is a template parameter and compiles away, this one cannot: its two error returns and the Annex B octal scan land in the parser body that every from_chars instantiation force-inlines, and fmt has to stay live across the integer scan.

This turned up while reviewing the v8.3.0 bump in ClickHouse (ClickHouse/ClickHouse#121383), which only ever passes chars_format::general. The bump alone cost about 3% on text float parsing.

The fix makes the rule a compile-time parameter of the parser body. parse_number_string keeps its signature and behaviour: it tests the format once and routes to a cold, never-inlined javascript instantiation, so callers that use it directly are unaffected. from_chars_float_advanced leaves for a cold from_chars_float_javascript as its first statement, before anything else is live, so it compiles to a tail call. Placement matters here: any later, and options stays live across the parse and the hot frame picks up two extra spill pairs, which costs as much as the original problem.

Measurements

Intel Xeon 6975P-C (Granite Rapids), clang 22.1.8.

ClickHouse, instructions in from_chars_caller<double>::call: 935 on v8.2.10, 1031 on v8.3.0, 944 on this branch.

ClickHouse, parsing a 20M row text column of Float64, perf stat for the whole run:

	instructions	cycles
v8.2.10	16.295G	3.179G
v8.3.0	16.437G	3.274G
this branch	16.354G	3.182G

benchmarks/benchmark.cpp, cycles per float, best of 7:

	canada.txt	mesh.txt
v8.2.10	47.03	24.78
v8.3.0	47.34	26.08
this branch	46.28	24.86

ctest is green with clang 22 and with gcc in C++20 mode. clang-format-17 --style=file reports no changes.

[lemire]

@Algunenano This looks good, running tests, and then I will merge.