[WIP]ggml-hexagon: create generalized functions for cpu side op

[chraac]

This PR refactors the Hexagon backend to generalize CPU-side operations and improve code maintainability. The primary goal is to reduce code duplication by merging similar operation handlers and standardizing buffer management.

Key Changes

1. Generalized Operation Handlers:

   • Merged ggml_hexagon_mul_mat into ggml_hexagon_binary.
   • Merged ggml_hexagon_mul_mat_id into ggml_hexagon_add_id (renamed to ggml_hexagon_binary_id).
   • These functions are now templated with <bool _IsSrc0Constant> to efficiently handle cases where the first operand (like weights) is constant, optimizing cache maintenance.

2. Refactored Buffer Management:

   • Introduced enum dsp_buffer_type to explicitly define buffer roles:
     • DSP_BUFFER_TYPE_DSP_WRITE_CPU_READ
     • DSP_BUFFER_TYPE_CPU_WRITE_DSP_READ
     • DSP_BUFFER_TYPE_CONSTANT

[chraac]

Here we check if the buffer is supported by this backend.

See also:
https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/ggml-cuda/ggml-cuda.cu#L3915

[chraac]

We can probably refactor these op functions into a single ggml_hexagon_generic function that handles null sources. This would reduce code duplication significantly. WDYT? @max-krasnyansky