fix: prevent integer overflow in candle backend sequence length calcu…

[xrwang8]

Fix integer overflow in Candle backend sequence length calculation

What does this PR do?

This PR fixes a critical integer overflow bug in the Candle backend that causes CUDA driver crashes and massive memory allocation requests (~18.4 EB) when processing certain batch configurations.

Root Cause: The issue occurs in backends/candle/src/lib.rs where sequence lengths are calculated using unsigned integer subtraction without overflow protection:

(batch.cumulative_seq_lengths[i + 1] - batch.cumulative_seq_lengths[i]) as usize

When cumulative_seq_lengths[i] > cumulative_seq_lengths[i + 1] (due to malformed batch data), the subtraction underflows, producing a very large u32 value that gets cast to usize, resulting in massive memory allocation requests that crash the CUDA driver.

Solution: Replace unsafe subtraction with checked_sub() to detect overflow conditions and fail fast with a clear error message:

batch.cumulative_seq_lengths[i + 1]
    .checked_sub(batch.cumulative_seq_lengths[i])
    .expect("Invalid cumulative sequence lengths: sequence lengths must be non-decreasing")
    as usize

Symptoms this fixes:

• CUDA driver errors: CUDA_ERROR_UNKNOWN
• HAMi OOM errors with extremely large memory requests
• Service becoming unresponsive after initial successful requests
• nvidia-smi showing corrupted memory usage displays

Error logs example:

[HAMI-core ERROR]: Device 0 OOM 18446744073702267944 / 24146608128
DriverError(CUDA_ERROR_UNKNOWN, "<Failure when calling cuGetErrorString()>")

Impact:

• Fixes critical CUDA driver crashes and OOM errors
• Improves error diagnostics and debugging experience
• Maintains full backward compatibility for valid inputs
• Negligible performance impact (single integer comparison per sequence)

Testing: Added unit test to verify the fix correctly detects and handles invalid cumulative sequence lengths.

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the documentation guidelines, and here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR.

@Narsil