Taking `requires_grad` propagation seriously in Thunder

Nobody has claimed this yet.

## 🚀 Feature

Add correct `requires_grad` attribute propagation in Thunder with OpInfo-based tests.

Currently, Thunder queries the `requires_grad` attribute only for compiled function inputs (nn.Module parameters, buffers, and args) and uses this attribute to cache compiled artifacts and mask the output of the compiled backward function.

We could also faithfully model the propagation of this attribute, as in PyTorch Eager, or consistently, according to the rules that we need to develop. However, the exact benefits are unclear and need to be collected to consider whether it's worth the effort.

If we decide to do the `requires_grad` propagation we would be able to faithfully mark some outputs of the compiled function with `requires_grad=False` by disconnecting them from the Autograd graph (https://github.com/Lightning-AI/lightning-thunder/issues/1733). The performance benefits of doing so remain to be evaluated.

cc @lantiga @mruberry

Start by reviewing the proposed OpInfo-based tests and comparing requires_grad propagation with PyTorch Eager. Collect the benefits and define consistent propagation rules before considering the output disconnection described in issue 1733. Done means the rules are agreed, implemented, and covered by OpInfo-based tests.

Written by the indexing model from the issue text.