Support raw dispatch in slangpy

[mkeshavaNV]

On discord we discusssed how to implement raw dispatches in slangpy.

slangtorch supports raw dispatches via launchRaw. slangpy currently supports this via 'mymodule.myfunc.dispatch', but it is not well maintained and should be retired

Here's the gist of the task here - Modify our existing calldata.py based path to be able to:

• take a thread count, in the event the kernel has a dimensionality of 0 (no arguments are vectorized) -
• able to directly call an entry point without wrapping it in a trampoline with a CallData structure if it is already tagged as a compute shader or cuda kernel
• able to generate a minimal wrapper for an entry point when the only thing missing is the compute shader / cuda kernel tag
• able to wrap a 'raw entry point' for the backwards pass, and detect that if the forwards function is marked [CUDAKernel], so should the backwards one be

[mkeshavaNV]

The first two tasks here come under slangtorch_parity. Rest are slangtorch_parity_polish