Add --precision flag for reduced-precision inference

[rioffe]

Selectively casts heavy encoder/backbone modules (monodepth_model, feature_model) to bfloat16 or float16 while keeping lightweight heads in float32 for numerical stability. Achieves ~2x inference speedup on MPS with bfloat16.

Summary

• Adds a --precision flag to sharp predict accepting float32 (default), bfloat16, and float16
• Only the heavy encoder/backbone modules (monodepth_model, feature_model) are cast to reduced precision; lightweight heads remain in float32 for numerical stability
• Forward hooks handle input/output casting so the rest of the pipeline is unaffected
• Documents the new flag in README with usage example

Test plan

• Run sharp predict without --precision flag and verify output is unchanged (float32 default)
• Run sharp predict --precision bfloat16 and verify Gaussians are produced without errors
• Run sharp predict --precision float16 and verify Gaussians are produced without errors
• Compare output quality between float32 and reduced-precision runs on a known input
• Verify inference speed improvement on supported hardware (CUDA/MPS) with bfloat16/float16