refactor(dataset-raw): bucketed dump partitions

[sistemd]

• In preparation for a Solana extractor refactoring, introduce fn bucket_size() -> Option<NonZeroU64> - a new method on the BlockStreamer trait.
• The return value of this method (if Some) is the size of the buckets in which this extractor fetches blocks. This matters for determining how to split block ranges for parallel streaming. For example, the Solana extractor would benefit from knowing that each instance of it is the only one responsible for buckets that encompass the range that the instance needs to extract.
• Refactors the dump logic to respect bucket_size (if Some) when partitioning block ranges that should be dumped.

[claude]

Review Summary

Critical

• BlockStreamerWithRetry does not delegate bucket_size(): The new bucketed partitioning path is unreachable in production because .with_retry() wraps the client before it reaches dump_ranges, and the retry wrapper inherits the default None return. This makes the Solana bucket_size implementation and split_and_partition_bucketed effectively dead code.

Bugs / Robustness

• Silent fallback in Solana bucket_size(): NonZeroU64::new(DEFAULT_SLOTS_PER_EPOCH) should use .expect() or a const assertion instead of silently returning None if the constant were ever 0.
• Potential arithmetic overflow in Phase 1: this_end + 1 and the bucket end calculation can overflow near u64::MAX (theoretical, but worth guarding).

Consistency / Guidelines

• Missing min_partition_blocks assertion: split_and_partition_bucketed lacks the invariant check that split_and_partition has for non-leftover partitions.
• Missing # Panics documentation: Both partition functions use assert! without documenting panic conditions per docs/code/rust-documentation.md.
• Documentation mismatch in extractors.md: bucket_size() shown as a required method (no body) but has a default impl; checklist says "four required methods" but there are now five.

Testing

• Additional test cases suggested: n=1, n > number of buckets, non-contiguous ranges with gaps across buckets, and a case exercising Phase 3 merge logic.