Improve the TIDL mechanism

[vsht]

Tensor decompositions of high rank integrals or integrals with many legs tend to become
very large. It is not efficient to load them from text files, as this may significantly
delay the loading of the package.

Instead, we should use the same mechanism as in Tarcer: The large text file is read
once and then converted into a machine specific binary format. Next time FeynCalc
is loaded, only the binary is picked up (which is fast).

Furthermore, we need to ship only a smaller subset of tensor decompositions
with the package itself. Very large decompositions should be made available via
a repository (e.g. as an add-on) so that people who need that can download them
separately.