Speed up tile reading

[clemsgrs]

Goal is2 tracks:

1- leverage cuCIM recently added support for batch reads (rapidsai/cucim#1007)
2- use super tiles to read larger regions that then get sliced in smaller tiles to reduce number of reads

First few benchmarking runs suggest:

1- cuCIM batch read alone already provides a ~3x speed up
2- using super tiles alone can yield a ~4x speed up (but depends on whether we can find super tiles, and if so, how many we find ; so results may vary from no speed up to even higher speed up depending on the slide geometry)
3- combining both naturally yields an even higher ~12x speed up

Mode	Tiles	Elapsed (s)	Read Calls	Tiles per Read	Tiles/s	MPix/s	Speedup vs wsd_baseline
wsd_baseline	15,950	1600.912	15,950	1.00	9.96	0.50	1.00x
wsd_supertiles	15,950	402.066	4,010	3.98	39.67	1.99	3.98x
cucim_batch	15,950	519.650	15,950	1.00	30.69	1.54	3.08x
cucim_batch_supertiles	15,950	128.114	4,010	3.98	124.50	6.25	12.50x