fix: llama-finetune backward pass crashes

[System64fumo]

Overview

Fixes multiple crashes in llama-finetune that prevent training.

Closes: #18499 #21037

Additional information

Note: This doesn't fully fix/improve/change the behavior of the finetune tool, It only fixes the crashes.
@JohannesGaessler I propose adding a message to let users know there are deeper issues with the tool. (As stated in #18499)

The changes were tested against a simple dataset just to confirm the crashes were fixed and no further major crashes were left.
Proper finetune testing was not done as i don't have the time to let this run against a massive dataset to see actual results, But a simple test to confirm basic functionality with llama-server/llama-cli was done and worked just fine.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES: I used GLM 5.0 Turbo to analyze the crashes and propose fixes then reviewed/picked a set of changes.

[JohannesGaessler]

Why are you making changes to the backend scheduler?

[System64fumo]

Why are you making changes to the backend scheduler?

Because otherwise:

Using host libthread_db library "/usr/lib/libthread_db.so.1".
0x0000ffff84ef40a8 in ?? () from /usr/lib/libc.so.6
#0  0x0000ffff84ef40a8 in ?? () from /usr/lib/libc.so.6
#1  0x0000ffff84ee6f34 in ?? () from /usr/lib/libc.so.6
#2  0x0000ffff84ee6f78 in ?? () from /usr/lib/libc.so.6
#3  0x0000ffff84f3e720 in wait4 () from /usr/lib/libc.so.6
#4  0x0000ffff85b05330 in ggml_print_backtrace () from /mnt/nas/llama.cpp/build/bin/libggml-base.so.0
#5  0x0000ffff85b054b0 in ggml_abort () from /mnt/nas/llama.cpp/build/bin/libggml-base.so.0
#6  0x0000ffff85b21468 in ggml_backend_sched_alloc_graph () from /mnt/nas/llama.cpp/build/bin/libggml-base.so.0
#7  0x0000ffff85b35c4c in ggml_opt_alloc () from /mnt/nas/llama.cpp/build/bin/libggml-base.so.0
#8  0x0000ffff899934b8 in llama_context::opt_epoch_iter(ggml_opt_dataset*, ggml_opt_result*, std::vector<int, std::allocator<int> > const&, std::vector<int, std::allocator<int> > const&, llama_batch&, void (*)(bool, ggml_opt_context*, ggml_opt_dataset*, ggml_opt_result*, long, long, long), bool, long, long, long) () from /mnt/nas/llama.cpp/build/bin/libllama.so.0
#9  0x0000ffff89993ac4 in llama_context::opt_epoch(ggml_opt_dataset*, ggml_opt_result*, ggml_opt_result*, long, void (*)(bool, ggml_opt_context*, ggml_opt_dataset*, ggml_opt_result*, long, long, long), void (*)(bool, ggml_opt_context*, ggml_opt_dataset*, ggml_opt_result*, long, long, long)) () from /mnt/nas/llama.cpp/build/bin/libllama.so.0
#10 0x0000aaaadaeb3684 in main ()

Unless you're suggesting this be fixed elsewhere?