Continue investigating LSTM speedups

[sffc]

Some things we think might work to make it faster still:

• Use compiler intrinsics for vectorization (Add configuration to use system BLAS in ML calculations #3214)
• Parallelize backward iteration
• Walk through the forward iteration with the string without storing it on the heap (Computing LSTM forward layer without allocating #3351)
• In-memory calculations (https://techxplore.com/news/2024-11-software-package-bypass-cpu-efficient.html)

[robertbastian]

Closing #3301 for now

[robertbastian]

Discussion: it should be possible to use simd intrinsics even from unaligned data, @sffc has some ARM code that maintains performance (please post)

[sffc]

Discussion: it should be possible to use simd intrinsics even from unaligned data, @sffc has some ARM code that maintains performance (please post)

#3430

[sffc]

Wonder if processing-in-memory (PIM) could improve the performance

https://techxplore.com/news/2024-11-software-package-bypass-cpu-efficient.html