`value_to_usize` loses precision

On 32-bits architectures, `usize` is 32-bits, by definition, so any conversion from a larger type into this will drop some bits. On 64-bits architecture, converting from 128-bits to 64-bits will also lose bits. It seems that it's only used for key paths computations, so for instance on 64-bits, that would require a static access in a list at an index >= 2**64 to cause an issue, so it's rather unlikely to happen, but it's still a real bug.

Search the repository for `value_to_usize` and trace its callers, especially the key-path computation code. Determine the intended behavior when a larger integer does not fit in `usize`, then inspect existing tests for a suitable regression case. Done means the conversion no longer silently loses precision on 32-bit or 64-bit architectures.

Written by the indexing model from the issue text.