Move more TPM modules and setup routines to common

[wdeconinck]

Another small step in direction of using multiple handles of different backend and resolution.

Moving more TPM modules and setup routines to a common precision/backend independent library means that there is no code duplication and/or multiple compilation and a possible reuse of the globals that are part of these modules across the different backends.

I managed to move all but one TPM modules.
The missing one for now is TPM_FLT because that essentially contains the Legendre coefficients for each backend in a specific precision. The CPU version also contains the fast-legendre coefficients and butterfly structures which are not ported to the GPU.
Because of this, it will be still some task to see how we can reuse the Legendre coefficients of same resolution across different precisions and backends without recomputing them.

Most work was spent in making TPM_FIELDS , and the flattened GPU arrays have been moved to a new TPM_FIELDS_FLAT module, and I understand that the code will be refactored soon to remove the use of the flattened arrays.

Note that results will not be bit-identical, but hopefully performance-neutral. Could someone benchmark/verify this for me?

[samhatfield]

Nice. As before, might take some time to look over this in detail.

[samhatfield]

Looks good, but I'd like a bit more assurance that the additional casting won't impact performance. Maybe it's insignificant, but we should demonstrate that with evidence. I'll try running some benchmarks.

[samhatfield]

Here are some performance results from --norms --niter 100 --nlev 137 --nfld 1 --vordiv --uvders --scders --meminfo:

commit: 9198e86

number_of_nodes	truncation	precision	inv_med	dir_med	loop_med	max_error_combined
1	502	sp	0.297	0.1541	0.4511	8.94e-06
2	633	sp	0.2606	0.1323	0.3929	4.47e-06
4	798	sp	0.2306	0.1255	0.356	3.46e-06
8	1006	sp	0.2303	0.1524	0.3826	6.07e-06

commit: d9ccd88

number_of_nodes	truncation	precision	inv_med	dir_med	loop_med	max_error_combined
1	502	sp	0.2963	0.1549	0.4511	6.66e-06
2	633	sp	0.2789	0.1413	0.4202	7.17e-06
4	798	sp	0.2306	0.1239	0.3544	4.72e-06
8	1006	sp	0.2143	0.1362	0.3509	7e-06

I think it's safe to merge this.

[marsdeno]

Looks good, and nice testing from @samhatfield