[inductor] Coordinate-descent tune compile-time combo kernel blocks (#189003)

### Summary

Enable runtime coordinate-descent tuning for combo kernels whose per-subkernel block sizes were selected by compile-time autotuning.
Compile-time autotuning stitches standalone winners into suffixed constexpr arguments such as `XBLOCK_0` and `XBLOCK_1`. Previously, this path did not register those fields with `CoordescTuner`, so coordinate descent could not refine the stitched block sizes.

This change:
- Registers suffixed block fields in largest-subkernel-first order.
- Propagates per-field minimum and maximum constraints.
- Honors TMA, `min_xblock`, and `min_rblock` requirements.
- Excludes body-baked and non-argument block fields.
- Bounds the optional all-directions Cartesian search for large combos.

#### Test plan
```bash
python test/inductor/test_coordinate_descent_tuner.py
python test/inductor/test_combo_kernels.py \
  -k test_combo_kernel_coordesc_tunes_largest_subkernel_first
```
The combo test covers both `combo_kernel_compile_time_autotune=False` and `True`.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/189003
Approved by: https://github.com/eellison