Reduce overhead in DataStream.submit_data().

[ehpor]

The old implementation performed type checking using Numpy functions, so that required a call from C++ back to Python, letting Python perform dtype operations, converting the result to a Python string, converting that string to a C++ string, then doing a C++ string comparison. The new code performs a C++ string comparison for the type directly using the dtype of the buffer format.

[ehpor]

Benchmarks: for 32x32 float32:

Macbook Pro M2 (Python 3.9.18):
develop: 4.93 µs ± 27.6 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
this PR: 674 ns ± 11.8 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

Hicatdeuxbis (Python 3.7.6):
develop: 13.9 µs ± 166 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
this PR: 1.73 µs ± 12.7 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

[ivalaginja]

What's with the addition of all the np.abs() in the tests?

[ivalaginja]

Nvm, your commit message explains it.

[ehpor]

@ivalaginja That's to avoid a warning when casting to unsigned integers from data that is both positive and negative. I presume it was there already, so it's kinda code sneak. It's not critical to this PR.

[ivalaginja]

Running the same benchmarks like @ehpor (after remembering that I had to recompile catkit2 😬 ), for 32x32 float32 data:

Macbook Pro M1 (Python 3.9.12)
develop: 9.6 µs ± 63 ns (mean ± std. dev. of 7 runs, 100,000 loops each)
tihs PR: 1.4 µs ± 5 ns (mean ± std. dev. of 7 runs, 1,000,000 loops each)

thd2-controller (Python 3.9.12)
develop: 14.5 µs ± 48 ns (mean ± std. dev. of 7 runs, 100,000 loops each)
this PR: 1.7 µs ± 12 ns (mean ± std. dev. of 7 runs, 1,000,000 loops each)

[ehpor]

Since this is a low-level change, I want to do a full end-to-end test on HiCAT before merging.

[ivalaginja]

Looks good to me.
Hold merging until Emiel gives green light from the HiCAT test.

[ehpor]

Benchmarks on HicatDeuxBis:

Develop:

This PR:

[ehpor]

Confirmed working on hardware.