Failing tests on newer CUDA hardware

Tried running the `candle-core` test suite on a 5070 and got these errors:

```
failures:

---- argmin_gpu stdout ----

thread 'argmin_gpu' panicked at candle-core/tests/tensor_tests.rs:548:5:
assertion `left == right` failed
left: [[[1], [0]], [[4294967295], [4294967295]]]
right: [[[1], [0]], [[1], [1]]]
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

---- min_gpu stdout ----

thread 'min_gpu' panicked at candle-core/tests/tensor_tests.rs:424:5:
assertion `left == right` failed
left: [[[0], [0]], [[0], [0]]]
right: [[[1], [1]], [[1], [2]]]

---- max_gpu stdout ----

thread 'max_gpu' panicked at candle-core/tests/tensor_tests.rs:506:5:
assertion `left == right` failed
left: [[3999, 3999]]
right: [[3998, 3999]]

---- argmax_gpu stdout ----

thread 'argmax_gpu' panicked at candle-core/tests/tensor_tests.rs:679:9:
assertion `left == right` failed
left: [[[4294967295, 4294967295, 189, 189], [189, 189, 189, 189], [189, 189, 189, 189], [189, 189, 189, 189], [189, 189, 189, 189]]]
right: [[[189, 189, 189, 189], [189, 189, 189, 189], [189, 189, 189, 189], [189, 189, 189, 189], [189, 189, 189, 189]]]

failures:
argmax_gpu
argmin_gpu
max_gpu
min_gpu
```

Start with candle-core/tests/tensor_tests.rs at the reported lines 424, 506, 548, and 679, then run the candle-core test suite on the newer CUDA hardware described in the issue. Investigate the min, max, argmin, and argmax GPU failures and identify the hardware- or CUDA-specific cause. Done means these four tests pass without changing their expected results.

Written by the indexing model from the issue text.