llama-fit-params: lower ctx size for multi GPU

[JohannesGaessler]

Fixes #18097 .

For multiple GPUs the context size reduction on master seems to be too optimistic:

// for multiple devices we need to be more conservative in terms of how much context we think can fit:
//   - for dense models only whole layers can be assigned to devices
//   - for MoE models only whole tensors can be assigned to devices, which we estimate to be <= 1/3 of a layer
//   - on average we expect a waste of 0.5 layers/tensors per device
//   - use slightly more than the expected average for nd devices to be safe

[verygreen]

This works, thank you very much.

The new suggestion is

-c 176384 -ngl 94 -ts 10,8,7,34,35 -ot blk\.9\.ffn_(gate|down).*=CUDA1,blk\.17\.ffn_(up|gate|down).*=CUDA2,blk\.58\.ffn_(gate|down).*=CUDA4

which fits with some space to spare, but I understand the VRAM waste is unavoidable because we cannot really split things on a byte boundary

[JohannesGaessler]

I didn't ping you yet because I wanted to rebase on top of the other fixes first but glad to hear it's working regardless.