Conditionally reduce max DRL ref MVs for speeds >= 2

[kslu-aom]

• Introduce the reduce_max_drl_refmvs speed feature to reduce the default DRL candidate count from 4 to 3 in higher encode speeds.
• Enable frame-level DRL adaptation (allow_frame_max_drl_bits = 1) when reduce_max_drl_refmvs is active: key, boosted, and low pyramid level frames (<= 2) retain 4 candidates (max_drl_bits = 3) to retain reference quality, while leaf frames use 3 candidates (max_drl_bits = 2).
• Condition newmv_drl_search_limit on reduce_max_drl_refmvs to ensure both ref_mv_idx 0 and 1 are searched before reusing results, preventing motion predictor diversity collapse.
• The existing --max-drl-refmvs CLI flag can override this behavior.

Evaluation based on 2abcfff with 33 frames at speed 2:

+---------+--------+--------+--------+--------+----------+----------+
| Summary |   Y    |   U    |   V    |  YUV   | Enc-time | Dec-time |
+---------+--------+--------+--------+--------+----------+----------+
| A1      | +0.05% | -0.05% | +0.10% | +0.05% |  98.1%   | 100.0%   |
| A2      | +0.08% | +0.27% | -0.03% | +0.08% |  97.5%   | 100.5%   |
+---------+--------+--------+--------+--------+----------+----------+

[kslu-aom]

The speedup-loss tradeoff is not good after rebase due to conflicts with recently added speed features. Marked as draft. I will rework this patch.

[yeqing-wu]

@kslu-aom, I’ve done this before, and at that time, the tradeoff wasn’t good either. However, it appears to be acceptable after the commit of this PR, but it’s not good again on the latest ToT.

[yunqingwang1]

@kslu-aom The code change affects speed 1 as well?

[kslu-aom]

@kslu-aom The code change affects speed 1 as well?

Some changes intended for speed 1 tests are included by mistake. I will update this PR later.

[kslu-aom]

@yunqingwang1 I have cleaned up the unintended changes. Please take another look

[yeqing-wu]

Overall LGTM, one minor suggestion to simplify the code