We should check and exclude data transfers from timing otherwise tests report unreliable values

[DiamonDinoia]

          Just a tiny text fix. I checked a cmake build on my local RTX-A6000.

But it has a weird timing inconsistency.
Eg, running it by hand:

(base) ahb@ccmlin019 /mnt/home/ahb/numerics/finufft/build> test/cuda/cufinufft1dspreadinterponly_test 1e7 1e8 1e-3 1e-2 f 0.0
spread-only test 1d:
	100000000 NU pts spread to 10000000 grid in 0.046 s 	2.17e+09 NU pts/s
	rel mass err 0.0002
interp-only test 1d:
	100000000 NU pts interp from 10000000 grid in 0.341 s 	2.93e+08 NU pts/s
	rel sup err 0.000497

It is nice to see the 2e9 NU pt/sec for spread, and this matches the expected from 1d1 actual NUFFT. But it is 10x slower for interp? Could you look into why that is?

I notice that 1d1 vs 1d2 are not different (but the setNUpts time is 10x longer than the exec): here I use precision d in order to test accuracy, but the speed is the same for f:

(base) ahb@ccmlin019 /mnt/home/ahb/numerics/finufft/build> test/cuda/cufinufft1d_test 1 1 1e7 1e8 1e-3 1e-2 d 0.0
[time  ] dummy warmup call to CUFFT	 0.00474 s
[time  ] cufinufft plan:		 0.0229 s
[time  ] cufinufft setNUpts:		 0.48 s
[time  ] cufinufft exec:		 0.0525 s
[time  ] cufinufft destroy:		 0.000309 s
[Method 1] 10000000 U pts to 100000000 NU pts in 0.556 s:      1.8e+08 NU pts/s
					(exec-only thoughput: 1.9e+09 NU pts/s)
[gpu   ] one mode: rel err in F[3700000] is 2.69e-05
(base) ahb@ccmlin019 /mnt/home/ahb/numerics/finufft/build> test/cuda/cufinufft1d_test 1 2 1e7 1e8 1e-3 1e-2 d 0.0
[time  ] dummy warmup call to CUFFT	 0.00385 s
[time  ] cufinufft plan:		 0.0225 s
[time  ] cufinufft setNUpts:		 0.48 s
[time  ] cufinufft exec:		 0.048 s
[time  ] cufinufft destroy:		 0.000335 s
[Method 1] 10000000 U pts to 100000000 NU pts in 0.551 s:      1.81e+08 NU pts/s
					(exec-only thoughput: 2.08e+09 NU pts/s)
[gpu   ] one targ: rel err in c[50000000] is 0.000339

These are all using method=1 as in your new tester.

It's very peculiar that setNUpts doesn't show up in the spread-only timing.
Could it be more cuda-event-synchronizations are needed?
Investigations welcome ! Thanks, Alex

Originally posted by @ahbarnett in #631 (review)

[ahbarnett]

The issue here is whether H<->D times are included in test/cufinufft{1,2,3}d_test
They should not be.

It could be because we switched from sync transfers to async, to allow users to increase throughput. This may mess up timings...