llama : reduce useless copies when saving session

[compilade]

Should help with #8915.

Some useless copies to immediately discarded temporary buffers were done as of #8699 because session size calculation and session file writing now share mostly the same code.

On CPU, the speed was reasonable, but on CUDA, as reported in #8915, this makes the session size calculation too slow.

To fix this, it's possible to simply avoid calling ggml_backend_tensor_get when the data won't be used (i.e. when calculating the session file size).

I've also eliminated the double tensor copies when saving the state to a buffer.

@josharian does this help with your use-case?

---

TODO

• Test state saving and restoring to and from a buffer
• Test session file saving and restoring
• Test sequence saving and restoring

---

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low

[josharian]

This looks great. Thank you!

[josharian]

OK, re-ran my profiling so I have real numbers to share.

This speeds up llama_state_seq_get_size by >25x, to the point that it is now legitimately cheap enough to use. :) Yay! Thank you!

This cuts roughly 5-6% off llama_state_seq_get_data, which is good, but not yet enough to make it usable in practice. I think batching up the tensor transfers (when contiguous) is probably the next thing to try out on this front.