Gridded PSF performance improvement

[keflavich]

This PR attempts to improve performance of GriddedPSF evaluation.

At a low level, some of these steps get ~10x speedups. However, I think they contribute negligibly to overall speed for most cases, so I'm not sure we should merge this - the slight cost in code complexity may be not worth it.

Evaluating this PR requires tests against a new benchmark:
astropy/photutils-benchmarks#5

Low-level improvement:

Test: 4x4 PSF grid, 51-pixel stamps, 2x oversampling, 10,000 evaluate calls

| Operation | Before (us) | After (us) | Speedup |
|-----------|------------|-----------|---------|
| `_find_bounding_points` | 159.52 | 13.11 | **12.17x** |
| `_calc_bilinear_weights + bounding` | 228.72 | 20.75 | **11.02x** |
| `evaluate()` full | 626.39 | 545.66 | **1.15x** |

If this 15% improvement really holds up on realistic cases, it's... maybe worthwhile?

Claude Haiku 4.5 contributed a bunch to this.