Cache the SiPixelFedCablingMapGPU across events

[fwyzard]

Try to restore both correctness and efficiency to SiPixelRawToClusterHeterogeneous.

Fails miserably with multiple GPUs...

[fwyzard]

This is less costly than #207:

Running 4 times over 4200 events with 1 jobs, each with 8 threads, 8 streams and 1 GPUs
  1759.7 ±   1.3 ev/s (4000 events)
  1755.4 ±   1.5 ev/s (4000 events)
  1749.2 ±   2.0 ev/s (4000 events)
  1763.4 ±   1.5 ev/s (4000 events)

Basically, it recovers the performance from #172 before the last commit 15c15ab , including the overall performance loss with respect to pre-#172:

Running 4 times over 4200 events with 1 jobs, each with 8 threads, 8 streams and 1 GPUs
  1797.6 ±   1.4 ev/s (4000 events)
  1797.8 ±   1.7 ev/s (4000 events)
  1801.5 ±   1.6 ev/s (4000 events)
  1800.8 ±   1.5 ev/s (4000 events)

[fwyzard]

With the validation on vinavx2 still broken, I have run a quick comparison by hand.

Summary table:

	reference-10824.5	development-10824.5	development-10824.8	testing-10824.8
Efficiency	0.4861	0.4852	0.2187	0.4841
Number of TrackingParticles (after cuts)	5666	5666	5666	5666
Number of matched TrackingParticles	2754	2749	1239	2743
Fake rate	0.0537	0.0537	0.0348	0.0358
Duplicate rate	0.0156	0.0151	0.0128	0.0154
Number of tracks	32379	32390	14566	31927
Number of true tracks	30641	30652	14059	30784
Number of fake tracks	1738	1738	507	1143
Number of pileup tracks	26878	26878	12375	26987
Number of duplicate tracks	504	488	187	491

Fakes and efficiency plots:

Looks like we recover the pre-#172 efficiency.

[fwyzard]

Superseded by #208 and #212