sw: Restructuring of BLAS and DNN kernels

[viv-eth]

This PR will introduce substantial changes to the SW library in the BLAS and DNN kernels.

GEMM

This PR comprises the following changes in the BLAS library:

1. We moved to a generic impl function that will fetch the correct kernel from the config.json file. All of this is wrapped inside a gemm_args_t struct that holds all kernel-related data. The arguments in the struct are loaded into the cluster's TCDM to avoid costly DRAM accesses when calling the function. The roadmap is to adapt all kernels in the sw library to this format to improve performance.
2. We added the option to run multiple tests in parallel.
3. We aligned the tests to work with the new GEMM layout and added a naive FP16 kernel for verification.
4. We added more assertions to the datagen script of unsupported or invalid configurations with the relevant error messages.
5. We adjusted other kernels, such as fused_concat_linear to the new layout.

FlashAttention-2

1. We added the transposition layer that is used inside the kernel.
2. We added the FP16 and FP8 MiniFloat implementations.
3. We added multiple configurations for testing.
4. We aligned the kernel to work with the new GEMM layout.
5. Several fixes to the kernel were integrated and casts from lower to higher precision and vice versa.

Treewide

This PR also includes some treewide changes.

1. Remove all math library related sources, which are no longer needed after hw: Fix memory consistency between int and FP datapaths #90.

2. Remove all uses of the $PYTHON environment variable, and instead consistently rely on shebang directives to pick up the active virtual environment's python executable.