Added support for overriding tensor buffer types

[zpin]

Equivalent to the -ot llama.cpp argument:

{"--override-tensor", "-ot"}, "<tensor name pattern>=<buffer type>,...",

Can be passed as an optionlal string to the Llama class using the new override_tensor parameter. Same format as the argument above.

Provides more control over how memory is used, letting you selectively place specific tensors on different devices, especially helpful when running large MOE models.