Reduce parser precompile workload

[quinnj]

Summary

• Narrow the PrecompileTools workload to core string, integer, Float64, and Bool parse/xparse cases.
• Stop precompiling BigFloat, Dates, SubString, and view inputs in the default workload.
• Add inference barriers around rare float widening branches so simple Float64 precompile avoids eagerly inferring the UInt128/BigInt overflow tree.

Measurements

Measured on Julia 1.12.5 in fresh temp projects/depots with dependency caches available:

• Original workload: ~4.96s Parsers precompile, 5.5 MB dylib, 1142 Julia symbols.
• This patch: ~2.54s Parsers precompile, 2.6 MB dylib, 388 Julia symbols.

First-call latency after precompile:

• parse(Int64, "123"): ~16.8 ms
• parse(Float64, "123.45"): ~18.3 ms
• parse(Bool, "true"): ~16.4 ms
• xparse(Float64, codeunits("123.45"), ...): ~2.0 ms
• xparse(Float64, Vector(...), ..., Any): ~2.0 ms

Tests

• julia --startup-file=no --project=. -e 'using Pkg; Pkg.instantiate(); Pkg.test(; coverage=false)'
• julia +1.6 --startup-file=no --project=. -e 'using Pkg; Pkg.instantiate(); using Parsers; @Assert Parsers.parse(Float64, "123.45") == 123.45; @Assert Parsers.xparse(Float64, Vector(codeunits("123.45")), 1, 6, Parsers.Options()).val == 123.45; println("julia ", VERSION, " focused checks ok")'
• git diff --check

Co-authored by Codex