Related: #9910

PR summary

This PR makes the fast path of the histogram significantly fast. The current benchmark vs main:

| Change   | Before [fefe4f5f] <main>   | After [fa366b8a] <hist-C>   |   Ratio | Benchmark (Parameter)                               |
|----------|----------------------------|-----------------------------|---------|-----------------------------------------------------|
| -        | 61.4±1μs                   | 40.4±0.7μs                  |    0.66 | bench_function_base.Histogram1D.time_small_coverage |
| -        | 883±40μs                   | 170±0.9μs                   |    0.19 | bench_function_base.Histogram1D.time_fine_binning   |
| -        | 861±40μs                   | 156±1μs                     |    0.18 | bench_function_base.Histogram1D.time_full_coverage  |

1. Introduces a optimization that were first discussed in ENH: push histogram calculations to compiled_base #9910
2. Fixes a (unknown?) bug where histogram would not go fast path for long double
3. adds new histogram files (which I am not sure if it is a good idea). this is mainly because it is possible to make histogram way more faster than this PR and histogram is a important function so I think it deserves its own file?

Other comments

I want to discuss this with someone experienced before I refine, implement the logic further into multidimensional case (for histogramdd). I tried to make the code look like other C files as much as possible (not sure if its smart thing to do)

I also feel like histogram deserves a couple of more benchmarks (especially for the slow path)

AI Disclosure

Claude was used as an AI tool.
Some parts (e.g template bodies) are written by me, some are recycled from the old python code, some are recycled from old PR (e.g error handling, numpy API etc) and some are written by AI (especially the weight type parts) and reviewed manually.
AI was also used extensively on Numpy API/Cpython specific parts, in parallel I have read some documentation and checked if I used them correctly, so be warned