[GGML][RPC] Support for models with non-512-aligned tensors over RPC.

[matt23654]

This PR adds support for quantized tensors with sizes not divisible by 512, such as those found in quantized versions of the Qwen2.5 72B model. Please also see discussion in #10943.

Changes

• Implemented forwarding to CUDA backend on server for init_tensor and get_alloc_size calls
• Forwarding all calls would result in unacceptable latency (tested and tok/s drops to 0.02). At the moment only calls for misaligned tensors are forwarded. There may be a better way of handling this in the future.

Performance Impact

• Qwen2.5 72B Q4_K_M: Coherent at ~4 tokens/s over GbE with mix of GPU/RPC/CPU. (previously outputted garbage)
• Existing models (e.g. LLaMA 3.3 70B Q4_K_M): Unaffected (~7 tokens/s over GbE with GPU/RPC)

Testing

• Perplexity validation with Tulu-3 8B:
  • Without RPC: 6.9362 ± 0.04526
  • With RPC: 6.9362 ± 0.04526 (identical results)
• test-backend-ops: Passed for supported backends
• Note: Unable to run full CI due to bandwidth limitations with model downloads

[rgerganov]

looks fine to me, some minor comments inline