[DRAFT] Optimize quality metrics performance

[alejoe91]

Depends on #4713

Optimizes speed and RAM usage (especially with lazy mode) for metrics computation by pre-extracting and sharing spike-vector arrays across metrics that need them using the tmp_data. This is especsially important in the context of lazy computation, as arrays need to be decompressed and loaded in memory.

Here is a benchmark with 20M spikes, amplitude_cutoff + amplitude_median + noise_cutoff computed together:

Backend	Lazy	Old compute time	New compute time	Speedup	Old peak RSS Δ	New peak RSS Δ
zarr	True	11.28 s	5.46 s	2.07x	1118 MB	1118 MB
zarr	False	2.63 s	2.31 s	1.14x	1197 MB	1198 MB
binary_folder	True	2.91 s	2.62 s	1.11x	1116 MB	1114 MB
binary_folder	False	2.62 s	2.14 s	1.22x	1193 MB	1194 MB

TODO:

• change unit/metric loop in compute for metrics that loop through unit to gather unit-level data
• test with fully lazy spike vector