perf: cache MiniMax H3 text conditioning

[xledx]

Summary

Cache MiniMax H3 text-only conditioning results so repeated requests with the same prompt can reuse the previously computed SDCondition.

The cache is intentionally conservative and is only enabled when:

• the model is MiniMax H3
• there are no effective LoRAs
• unconditional conditioning is not requested
• there are no image or MiniMax H3 presentation references

Non-cacheable H3 requests invalidate the cached condition.

Related Issue / Discussion

None.

Additional Information

Validated with MiniMax H3 on an Intel Mac, CPU-only, using sd-server so consecutive requests shared the same model context.

On a repeated identical text-only request, the runtime log reported:

H3 conditioning cache hit
get_learned_condition completed, taking 0.00s

[xledx]

Retested on the latest master.

• Confirmed the conditioning cache hit on an Intel Mac, CPU-only.
• get_learned_condition: 199.97s → 0.00s
• Wall-clock generation time: 6m36s → 2m59s
• Same prompt, seed, and generation settings were used for both runs.

The patch also cherry-picks cleanly onto the latest master, and both sd-cli and sd-server build successfully.