Improve Synet API documentation

[ermig1979]

Summary

• Improve SimdLib.h descriptions for Synet add, bias, channel-sum, 8-bit conversion, FP32/BF16/INT8 convolution, and FP32 deconvolution APIs based on their implementation behavior.
• Document per-channel formulas, NCHW/NHWC offsets, BF16 conversion, supported Add16b types/shapes, UINT8 saturation ranges, convolution/deconvolution context validation, buffer sizing units, parameter and statistics handling, implementation info strings, quantization behavior, and forward behavior.

Testing

• git diff --check
• git diff --name-only
• git status --short --branch

Full test suite was not run per request.

  

[Copilot]

Pull request overview

This pull request updates the public Synet API documentation in SimdLib.h to more accurately reflect the actual runtime behavior and constraints of several Synet operations (add/bias, channel sum, 8-bit conversion, and convolution/deconvolution contexts).

Changes:

• Expanded API docs with explicit per-format (NCHW/NHWC) indexing formulas and BF16/FP32 conversion semantics.
• Clarified convolution/deconvolution context validation rules (type/format constraints) and output shape formulas.
• Documented buffer sizing units (elements vs bytes) and clarified implementation info/parameter attachment behavior for convolution contexts.