fix: properly free BeamState on error paths to prevent segfaults (closes #675)

[zclawz]

Problem

Fixes #675 — zeam segfaults on InvalidPostState when processing a cached block.

Root cause: In onBlock's computedstate: labeled block, cpost_state is heap-allocated with allocator.create and then deep-cloned via sszClone (which allocates interior ArrayList fields like historical_block_hashes, justified_slots, etc.). If verifySignatures or apply_transition subsequently fails, there were no errdefer cleanup paths — the partially-mutated state leaked, causing UB and the observed segfault.

Specifically: apply_transition calls process_slots which appends to historical_block_hashes before the post-state root check. When InvalidPostState is returned, those appended bytes are never freed → dangling allocations → segfault.

Fix — three errdefer additions in chain.zig

1. onBlock — inside computedstate: block

LIFO errdefer pair: destroy registered first (covers sszClone failure), then deinit registered after clone succeeds (covers verifySignatures / apply_transition failure). When both fire, deinit runs first freeing interior fields, then destroy frees the outer struct.

2. onBlock — function level (post-state ownership)

If computedstate: succeeds but a later step fails (forkChoice.onBlock, updateHead, InvalidSignatureGroups), post_state was previously leaked. Added errdefer guarded by post_state_owned flag.

3. BeamChain.init — cloned_anchor_state

Same create+sszClone pattern without cleanup if states.put fails.