Optimize loop hafnian: share powertrace computation between f_loop calls

[seba2390]

Context:
In _calc_loop_hafnian_batch_gamma_even and _calc_loop_hafnian_batch_gamma_odd, the functions f_loop() and f_loop_odd() are called inside a loop over displacement vectors (k in range(n_D)). Both functions internally compute powertrace(AX), however, the matrix AX depends only on the edge configuration (j), and not on the displacement vector (D[k]). The original code redundantly computes the same powertrace 2 × n_D times per j iteration.

Description of the Change:

1. Modified f_loop and f_loop_odd in _hafnian.py to accept powtrace_arr as a parameter instead of computing it internally.
2. Updated callers in loop_hafnian_batch.py and loop_hafnian_batch_gamma.py to compute powertrace once per j iteration and pass it to both functions.

Benefits:

• 2-6x speedup on loop_hafnian_batch_gamma for typical batch sizes (10-50 displacement vectors)
• Speedup scales with batch size — minimal overhead for batch_size=1, significant gains for larger batches
• Benchmark results on Apple M3 Pro and M1 Pro available at: https://github.com/seba2390/thewalrus_opt_benchmark

Possible Drawbacks:

• Changes the internal API of f_loop and f_loop_odd (now require powtrace_arr parameter). These are internal functions not exposed in the public API.

Related GitHub Issues:

None