Decouple floating-point loop unrolling from DSP extensions

[sylvesterkaczmarek]

Summary

• select the loop-unrolled scalar implementations of arm_conv_f32, arm_conv_partial_f32, arm_correlate_f32, and arm_correlate_f16 using ARM_MATH_LOOPUNROLL
• retain the existing NEON selection for the f32 functions and the autovectorisation fallback for f16
• update the closing preprocessor comments to describe the actual conditions

Context

ARM_MATH_DSP represents integer DSP extensions and does not determine whether these floating-point implementations can use manual loop unrolling. The existing outer guards therefore prevent ARM_MATH_LOOPUNROLL from selecting the unrolled scalar path unless ARM_MATH_DSP is also defined.

This change makes the outer selection consistent with the loop-unroll option discussed in #193 while leaving the algorithms unchanged.

Fixes #193.

Validation

• git diff --check
• compiled all four affected translation units with GCC 13.3.0 and -Wall -Wextra -Werror -Wmissing-prototypes under:
  • ARM_MATH_LOOPUNROLL
  • plain scalar configuration
  • ARM_MATH_DSP without loop unrolling
  • ARM_MATH_LOOPUNROLL with ARM_MATH_AUTOVECTORIZE
• linked the patched unrolled and scalar implementations under separate symbols and compared representative results for f32 convolution, f32 partial convolution, f32 correlation, and f16 correlation; all outputs matched within the selected floating-point tolerances

The CMake-based repository suite was not run because CMake is unavailable in the local environment.