cpu: use templates for de-duplicating some operators

[cmdr2]

Note: This PR only deletes lines from ggml-cpu.c, it does not modify any functions in it. I'm not sure why git diff tries to combine them as changes. Standard diff shows the correct output. Actual diff of ggml-cpu.c: https://gist.github.com/cmdr2/a76df5af311417619788e8330b1908b3

This PR de-duplicates some of the easily-templatized functions in ggml-cpu.c. It takes inspiration from binbcast.cu.

• Binary: add, sub, mul, div
• Unary: abs, sgn, neg, step, tanh, elu, relu, sigmoid, hardsigmoid, exp, hardswish, sqr, sqrt, sin, cos, log

This removes the op implementation functions from ggml-cpu.c (around 2000 lines). As a side-effect, all the functions now support bf16 as well as non-contiguous src1.

The next PRs will attempt to:

1. Use the traits table to remove the ugly-looking mass of template parameters and if-else conditions. That will enable cleaner support of quantized operations. For e.g. taking inspiration from the quantized add function.
2. Use row-based type conversions, to work with the existing traits table functions.
3. Move even more operator functions to C++ files.

The performance is the same as the current implementation. It also passes all the runners on ggml-ci, which tested non-contiguous inputs (in SAM) and vDSP on Mac.

[cmdr2]

I'm happy to modify the approach based on feedback. This was one way to get this de-duplicated, without changing too many things at once.

[cmdr2]

Also, please let me know if the templates-based approach is too ugly. I can fast-track the traits table-based approach (my next PR), and use row-wise conversion functions instead.

I was trying to minimize the amount of changes in this PR, which is why I left that out. The existing implementation doesn't use row-wise conversions either, so this PR just refactors that.

[cmdr2]

I have a cleaner solution coming up shortly, please hold off.