ggml : extend the GGML_SCHED_NO_REALLOC debug logic of the scheduler

[ggerganov]

cont #17276

Graph reallocations by the backend scheduler can be expected in various cases when the graph topology becomes different from the one that was used initially after constructing the scheduler. For example, the scheduler of llama_context will likely reallocate when:

• Changing LoRAs
• Switching from token to embedding input in llama_batch
• Attaching different set of samplers (sampling : add support for backend sampling #17004)
• etc.

The less expected cases are the ones similar to the case in #17143 where it's more difficult to predict that a reallocation would occur since the graph topology remains the same.

The GGML_SCHED_NO_REALLOC macro is now targeted towards detecting the "unexpected" reallocations.

Also, can now override the reallocation debug behavior via a new environment variable GGML_SCHED_DEBUG_REALLOC:

# abort only on unexpected reallocations (i.e. same as building with -DGGML_SCHED_NO_REALLOC=ON)
GGML_SCHED_DEBUG_REALLOC=1 ...

# abort on all reallocations
GGML_SCHED_DEBUG_REALLOC=2 ...