dont (immedietly) promote fa-auto -> fa-on

[JorisBolsens]

Overview

in #25871 if FA was set to AUTO and ctv is quantized, then it automatically sets FA to enabled. That causes the check on src/llama-context.cpp:550 to never run, and causes the guard on src/llama-context.cpp:459 to never trip. Which means flash attention automatically offloads to CPU in case it is not supported in backend (often when ctk != ctv)
There is no log line showing this is happening, which means performance is quietly degraded, and debugging is very difficult.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES
  • AI was used to identify why performance of the model was (much) lower than expected, and was used to identify the code responsible for the issue.

[HermiG]

This was a gotcha for me when I didn't realize my llama-server was compiled without FA_ALL_QUANTS. Silently offloading to CPU should not be a default behavior; such a warning would have saved a lot of confusion.