Automatize the reuse of the CUDA stream of an input product

[makortel]

This PR addresses #278. Basic idea is that the first CUDAScopedContext constructor reading a CUDAProduct<T> object will re-use the CUDA stream of that product, and all the subsequent ones will create a new CUDA stream.

As an example, let's say we have the following modules/products

   A
 / | \
B  C  D
      |
      E

The product A is produced in one CUDA stream. One of the B, C, or D will re-use the same CUDA stream, the other two will create new ones. The E will use the same CUDA stream as D.

I believe this approach is the best we can do in a simple way using only local information.

[fwyzard]

Interesting approach, let's run the usual tests.

[fwyzard]

No impact on the throughput.

Before: 1846.2 ± 15.6 ev/s
After: 1843.0 ± 9.3 ev/s

[fwyzard]

No impact on the tracking performance:

	development-10824.52	testing-10824.52
Efficiency	0.5801	0.5801
Number of TrackingParticles (after cuts)	5575	5575
Number of matched TrackingParticles	3234	3234
Fake rate	0.0342	0.0342
Duplicate rate	0.0004	0.0004
Number of tracks	44607	44607
Number of true tracks	43080	43080
Number of fake tracks	1527	1527
Number of pileup tracks	38249	38249
Number of duplicate tracks	16	16

[fwyzard]