Checking for blas flags has a high import cost on the library

[ricardoV94]

Description

On my machine, which uses openblas, the check_blas, ultimately triggered in these lines when pytensor is imported:

pytensor/pytensor/tensor/blas_headers.py

Lines 1006 to 1007 in 55b2f4f

	if not config.blas__ldflags:
	_logger.warning("Using NumPy C-API based implementation for BLAS functions.")

Takes 2s out of the total 3s import time. The rest is due to scipy imports:

[ricardoV94]

At the very least we should explain in the docs how to store the computed flags in pytensorrc to avoid doing it after the first time