When we set reverse propagate_in_video, a data type mismatch error occurs.

**Background**
`propagate_in_video` allows passing `start_frame_idx` to specify the starting frame for propagation, `max_frame_num_to_track` to set the maximum length of propagation, and `reverse` to control the propagation direction, where `True` indicates reverse propagation.

**Error**
When we set `reverse=True` propagation from the most recent frame for a fixed length. We got an error：
`RuntimeError: mat1 and mat2 must have the same dtype, but got BFloat16 and Float`

**Reason**
When setting reverse propagation from the most recent frame for a fixed length, the `_prepare_memory_conditioned_features` method in the `SAM2Base` class will enter the if not `is_init_cond_frame:` branch.
This branch is generally not entered during forward propagation because the first frame in forward propagation is usually a condition frame. The purpose of this branch is to condition the visual features of the current frame on previous memory.
In the aforementioned branch, when `feats = prev["maskmem_features"].to(device, non_blocking=True)` is retrieved, the value is `torch.bfloat16` , which can cause subsequent `MemoryAttention` data type mismatch errors.

**Temporary solution**
add `feats = feats.to(torch.float32)`

**Question**
Why `prev["maskmem_features"]` in `SAM2Base` class is `torch.bfloat16` dtype？Where is it generated?

Start in the SAM2Base._prepare_memory_conditioned_features method and reproduce reverse propagation from the most recent frame with a fixed length. Trace where prev["maskmem_features"] is generated and confirm its dtype before MemoryAttention receives it. Done means the reported reverse-propagation path no longer raises the BFloat16/Float mismatch and the relevant behavior is covered by a regression test.

Written by the indexing model from the issue text.