Optimize IK solver performance (1.3-1.6x speedup)

[Daoming-Chen]

Summary

This PR optimizes the PseudoInverseIK solver performance through several targeted improvements:

• Eliminate redundant FK computation: The calc_jacobian function was computing forward kinematics twice per iteration. Now reuses the accumulated transform from the Jacobian loop.
• Cache regularization matrix: Pre-compute reg = regularization * torch.eye(6) once in __init__ instead of every iteration.
• Add torch.compile support: Optional JIT compilation for PyTorch 2.0+ via use_compile=True parameter.
• Reduce memory allocations: Pre-allocate delta pose buffer and use in-place tensor operations where safe.

Performance Results

Benchmarked on UR5 robot with 500 goals, 10 retries, 30 max iterations:

Device	Before	After	Speedup
CPU	472.43 ms	354.35 ms	1.33x faster
CUDA	172.30 ms	104.88 ms	1.64x faster

• No regression in convergence rate (~79% on CPU, ~78% on CUDA)
• No change in solution quality

API Changes

New optional parameter for PseudoInverseIK:

ik = pk.PseudoInverseIK(chain, use_compile=True, ...)  # Enable torch.compile (PyTorch 2.0+)

Test plan

• All existing tests pass (pytest tests/)
• Benchmarked on CPU and CUDA
• Verified convergence rate unchanged
• Verified solution quality unchanged (< 1e-6 difference)