Butterfly warp-reduction for packet scatter-reduce

[lnuic]

Adds scatter_reduce_packet_dynamic, the runtime-N analogue of gather_packet_dynamic.

Bumps ext/drjit-core for the butterfly packet path.

[wjakob]

Some (unordered) feedback:

I think that the packet atomics in OptiX aren't supported in all driver versions.
I was surprised that the guards just gate on "ts->compute_capability >= 70" -- are you sure that's correct?

Some of the comments feel a bit AI-generated, e.g. the "amortizing .. scaffolding" and "mirrors" .. are mentioned several times. One comment even mentions pixels in the Dr.Jit-Core PTX generation:

// Butterfly warp-reduction over the whole packet collapses up to 32
// lanes' contributions targeting the same pixel down to a single set

[wjakob]

(Updated the comment). Note that this is actually mostly about the Dr.Jit-Core parts.

[lnuic]

The cc >= 70 / ptx >= 63 gate is for match.any.sync (used for pre-aggregation), not for the vector atomics.
Vector atomics use the same supports_vector_reduction check as before; if they're unavailable the butterfly still runs and the leader just emits N scalar atomics.
Also cleaned up the comments.