Added whisper state + default state on the whisper_context

[sandrohanea]

In order to reuse memory and run multiple transformations on a single model, added the whisper_state (which is keeping the state of the current transformation).

In order to not be a breaking change and keep the usage of whisper_context to minimum, I added a default_state on the context and the current methods (e.g. whisper_full) will use that context in order to run the transformation.

Ofc, these methods are not thread safe, but additional methdos are created where the state can be externally provided (e.g. whisper_full_with_state )

For the callbacks, the state is also provided so bindings need to be updated for this, but the other methods are still valid.

[sandrohanea]

Hey @ggerganov,
This PR is an alternative to #494 where the existing functionaliy is kept the same (by using a default_state) and additional capability (to run multiple transformations on the same context) is provided with newer functions.

Let me know what you think about this simpler approach.

If you'll like it, we can create some follow-up tasks to update emscripten + various samples to not create so many contexts anymore and save some memory. :)

[ggerganov]

Nice - well done!

We need to update the callbacks in the examples.

If you'll like it, we can create some follow-up tasks to update emscripten + various samples to not create so many contexts anymore and save some memory. :)

Hmm, where do we create many contexts?

[sandrohanea]

Nice - well done!

We need to update the callbacks in the examples.

If you'll like it, we can create some follow-up tasks to update emscripten + various samples to not create so many contexts anymore and save some memory. :)

Hmm, where do we create many contexts?

Here for example: https://github.com/ggerganov/whisper.cpp/blob/master/examples/stream.wasm/emscripten.cpp#L16

But there are other samples which are creating multiple contexts like that.

[ggerganov]

@sandrohanea
Ah, this looks like it creates multiple contexts, but in reality it is just one.
Initially I was thinking the WASM lib should support multiple contexts, so I created this global map. And then it propagated in various examples. But it is actually not needed. I should remove it and replace it with a single instance.