fix(colgrep): persist FP32 override so settings --fp32 forces full precision (#130)

[raphaelsty]

Summary

Fixes #130. colgrep settings --fp32 was a no-op on non-CUDA builds: it printed success but the index/search path kept selecting model_int8.onnx.

Root cause: the --fp32 branch called config.clear_fp32() (sets the field to None). Config::use_fp32() resolves a missing value per build:

#[cfg(feature = "_cuda")]    { self.fp32.unwrap_or(true) }   // FP32 default
#[cfg(not(feature = "_cuda"))] { self.fp32.unwrap_or(false) } // INT8 default

So on CoreML/CPU builds, clearing → None → use_fp32() == false → INT8. Clearing is not equivalent to forcing FP32.

Changes

• --fp32 now persists Some(true) (set_fp32(true)) so use_fp32() returns true on every build → forces model.onnx.
• New --default-precision flag explicitly resets to the build default (FP32 on CUDA, INT8 otherwise). This is what clear_fp32() now backs, mirroring the existing reset-to-default pattern (e.g. --pool-factor 0, --clear-ignore).
• Status output shows the effective precision and whether it's an explicit override or the build default (previously hard-coded fp32 (default), which was itself misleading).
• Help text + README updated.

Tests

• test_set_fp32_forces_fp32_regression_130 — asserts set_fp32(true) ⇒ use_fp32() == true, set_fp32(false) ⇒ false, clear_fp32() ⇒ None (the regression).
• test_fp32_override_survives_serialization — JSON round-trip keeps Some(true) and use_fp32().
• Manual end-to-end (isolated config dir):
  • settings --fp32 → precision: fp32, persists "fp32": true
  • settings --default-precision → precision: int8 (build default) on a non-CUDA build
  • settings --fp32 --int8 → conflict error

make ci-quick (fmt + clippy -D warnings + tests) passes via the pre-commit hook.

Closes #130