server: reset counter related to kill-switch on client error

[SoftwareRenderer]

This avoids inadvertently triggering a server kill switch.

If the client sends a request that exceeds the configured context size, an appropriate HTTP 400 response is provided and no tokens are generated.

However since no tokens are generated, update_slots() increments n_empty_consecutive. If the client sends 4 such messages in a row, the server terminates.

This change resets the counter when the client's request exceeds --ctx-size. While I was looking at this, neighboring code had similar patterns relating to the same ERROR_TYPE_EXCEED_CONTEXT_SIZE, however I'm not familiar with those other conditions.

Steps to reproduce issue:

1. Start llama-server with a tiny context e.g. --ctx-size 10
2. Send text that exceeds the context size 4 times in a row

[ggerganov]

It would be better to move this to launch_slot_with_task() - right before returning true.

[SoftwareRenderer]

It would be better to move this to launch_slot_with_task() - right before returning true.

Done. I was worried that it might interfere with the debugging efforts in #20277 .

After looking at that issue closer, if I'm understanding it correctly: the loop occurs within the same task and slot. So resetting the counter won't impact that because the reset occurs before a new task/slot.