Improve performance for endian read/write functions

[armeenm]

Functions which accept endianness as a template parameter can benefit from constexpr branching to determine if byteswapping should take place or not. I benchmarked using the writelog example. Before:

500k double append (test2.log): 44470
50 string append (test-string.log): 262
Double array append (test-double-array.log): 6223
String array append (test-string-array.log): 330

After:

500k double append (test2.log): 38896
50 string append (test-string.log): 257
Double array append (test-double-array.log): 6170
String array append (test-string-array.log): 257

These results are fairly consistent across runs (generally a ~6ms improvement for the 500k double append).

[PeterJohnson]

Please separately PR the endian templated byte_swap change to the main wpilibsuite repo, since it could be committed independently.

Regarding the macros, we'll need to see if this fails when run against CI, in particular on Mac.

[PeterJohnson]

The other thing that would be great to have is real unit tests (e.g. using GoogleTest). We might need to do some work to enable use of temporary filenames though.

[armeenm]

I just realized the endian stuff is from LLVM. Do we still want to make changes to that?

[PeterJohnson]

I'm okay with making a change there; we don't want to do dramatic things like reformatting code, but minor changes I can deal with when merging from upstream 1-2 times a year.

[PeterJohnson]

It might be good to PR that upstream as well though.