llama-bench: fix accumulated load_time in perf timings

[abhinavuser]

Overview

Fix accumulated load_time in llama_perf_context_print when running llama-bench with multiple parameter sets (e.g. -n 4,8,16,32).

The load_time keeps growing because each new context inherits t_start_us from the model's original load timestamp, but the model is reused across runs. Added llama_perf_context_reset(ctx) after context creation to reset the timing baseline per iteration.
Fixes #9286

Additional information

Spotted this while benchmarking with verbose output — the load_time went from ~1s to ~33s across 7 runs even though the model was only loaded once.

Requirements

• I have read and agree with the contributing guidelines(https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md)
• AI usage disclosure: No

[abhinavuser]

not AI generated, i traced the bug through the code, each new context copies t_start_us from the model's original load timestamp (llama-context.cpp line 35), so when the model is reused across bench iterations the load_time just keeps growing
the fix is just resetting the perf context so t_start_us gets freshened