Add reproducible profile-guided builds and benchmarks

[Matei02355]

Provide an optional native-host PGO build workflow so developers can train and measure bat on their own files.

The Python 3 script builds with the matching toolchain's LLVM tools, gathers fresh profiles, merges them, and rebuilds with profile-use. It preserves compiler flags, isolates each run, supports paths containing spaces, and verifies output hashes before reporting success. An optional baseline build measures alternating, warmed runs and records raw timings, medians, compiler details and output hashes in JSON. Documentation explains setup, reproducibility and workload-dependent results. Normal release builds remain unchanged; BOLT and release-binary training are outside this change.

Addresses #2701.

Validation: completed all three release builds with Rust 1.91.1 / LLVM 21.1.2 on x86_64 Linux, offline, in a work directory containing spaces. Used two training repetitions and ten benchmark repetitions. Baseline, instrumented and optimized output hashes matched for all six training file/mode combinations. Separate large JSON and Python files not used for training also produced identical baseline/optimized output in both modes.

Measured medians (milliseconds; this machine/run only):

Input	Mode	Baseline	PGO	Baseline/PGO
src/printer.rs	highlighted	47.450	44.364	1.070
Cargo.toml	highlighted	10.259	8.891	1.154
README.md	highlighted	93.237	86.983	1.072
src/printer.rs	plain	2.076	1.936	1.072
Cargo.toml	plain	1.705	1.594	1.070
README.md	plain	2.174	2.252	0.965
JSON holdout	highlighted	1100.787	982.980	1.120
Python holdout	highlighted	1052.670	990.963	1.062
JSON holdout	plain	15.375	14.838	1.036
Python holdout	plain	7.780	7.436	1.046

Small timing differences can be noise; the plain README case regressed slightly. The holdout JSON contains 10,000 indented objects with id/name/enabled/values fields; Python repeats a three-line function 15,000 times. Compiler-flag inheritance and invalid inputs were also checked.