Allow graceful failover when one tunable panics

[wingertge]

Catches panics that occur during the benchmarking phase of autotuning and gracefully falls back to the next algorithm in line.

This allows selecting algorithms based on things like memory constraints (without adding Result types all through the memory allocation chain) or even things like potentially unsupported GPU features. Note that this only catches panics during benchmarking, not during regular execution. If an algorithm fails at that point it will still give error feedback to the user.
If all tunables panic, propagates the first panic.

One thing worth mentioning is that while it will gracefully continue to the next algorithm, it will not necessarily continue quietly. The error message still gets printed to the console, which could confuse users. Unfortunately I don't think that can be prevented due to how panic is implemented, the print gets output before unwind gets to the benchmark code where the unwind is caught. This can be prevented with a custom panic hook, but that would affect all panics everywhere, which is a bit too intrusive.

Testing

All existing tuners work as expected with all tests passing, and the fallback works for im2col convolution when called with parameters that cause an out of memory error.

[nathanielsimard]

Thanks for the contribution, and yes I think it makes a ton of sense to have this into autotune! I only have one comment.