Add build instruction of native windows build and fix ROCM7 build issue

[mmis1000]

This Pr includes

• Build instructions of native windows build
• Quick build script to compile the whole project in clean windows environment
• Fix of CMAKELIST related to bad WIN32 assumption
• Fix of CMAKELIST related to intel DNNL layout change on newer version from the official site
• Fix of ROCm 7 related name change and include change
• Fix of non hip-enabled build
• Workaround of graphic card crash on windows Windows benchmark crash on 9070xt + rocm7 #14

This Pr DOES NOT includes

• New ROCm 7 bf16 type ❌

[mega-ice]

I successfully compiled CTranslate2-rocm and passed the tests (not int8) on ROCm 7.0.2 GFX1101.

I have just updated the definitions in /src/cuda/primitives.cu based on this merge request.

then i was able to compile and run faster-whisper, whisperx (with torch-2.8.0+rocm7.0.2 stack)

[Bazza-63]

Thanks for this.

I can compile it successfully, completing the test run fine (bar the int8 but that's always been broken). The faster-whisper works too.

The thing I can't figure out is the dynamic library linking. If I explicitly add the intel libraries to path then the first test stalls taking 44 seconds on my 7800 XT. Othertimes it'll work and get within a few seconds of WSL2 but then never work again (even with system wide linkage).

Anyway thanks for this.