Rewrite f32.exp and f32.tanh?

[makotokato]

lstm uses f32.exp and f32.tanh, so we depends on num_trait with std. Should we remove num_trait and std dependency, then rewrite f32.exp and f32.tanh?

[aethanyc]

cc @sffc

[sffc]

CC @younies

[robertbastian]

num_traits seems to be unused currently. However it can provide f32::exp and f32::tanh (through its Float extension trait) in #[no_std] if we enable the libm feature. Alternatively we could use libm directly, but it's less ergonomic and num_traits is a transitive dependency already.

[robertbastian]

I believe the no-stdness of LSTM was the reason it's still behind a feature and not included in the meta crate. Should we include it now?

[sffc]

I believe the no-stdness of LSTM was the reason it's still behind a feature and not included in the meta crate. Should we include it now?

It's one of several blockers; see the meta-issue #2259

[robertbastian]

Those are blockers for stabilizing the whole crate. I'm talking about LSTM, which does not need to be behind a feature anymore (before it had to be because otherwise icu with experimental features wouldn't have been no_std). Unless we want to keep it behind a feature even after stabilization?

[sffc]

The LSTM features are kind-of a mess right now. Enabling LSTM disables Dictionary. That's one of the things that needs to be fixed before stabilizing.