Speed up template similarity with pair-level parallelism

[JESUSROYETH]

Template similarity currently parallelizes over source rows. A common soft merge compares one new template with hundreds of targets, but exposes only one parallel task.

This flattens connected (source, target, active_channels) pairs into typed lists and runs prange over them. For a square matrix, row i has N - i pairs, so pair-level scheduling also balances work better than complete rows. Support selection, cosine/L1/L2 arithmetic, lag symmetry and disconnected-pair behavior stay unchanged.

I ran five paired fresh-process workflows after one warm-up on a GCP c3-standard-8 (Xeon Platinum 8481C, 8 vCPUs), with 256 units, 384 channels and 20 public soft merges.

Workflow	Baseline median (range)	Candidate median (range)	Change
Complete setup + 20 soft merges	17.767 s (17.323-17.844)	12.771 s (12.682-12.848)	-28.1%
Initial template similarity (fresh process, includes JIT)	11.585 s (11.116-11.634)	9.644 s (9.574-9.699)	-16.8%
Median soft merge	227.2 ms (224.8-229.3)	73.6 ms (73.5-73.6)	-67.6%

The complete workflow includes imports, analyzer setup, template calculation and JIT compilation. Every paired run improved by 26.3-28.4% and returned the same checksum.

On the public 10-second MEARec fixture, five warmed soft merges went from 6.17 ms to 2.42 ms, with exactly equal complete matrices. A separate 300-case sweep matched NumPy across all metrics and support modes, sparsity and varied shifts. The complete template-similarity module has 19 passing tests, the dense/sparse extension merge test has 2 passing cases, and Black is clean.

Related to #4310.

[yger]

This looks good to me, this is a nice observations and since the speedup seems to be there, good catch!

[JESUSROYETH]

Thanks :)