Fix split-buffer remap state consistency

[cursor]

Bug and impact

• Placeholder channels remapped by the first publisher replaced their CCB/BCB shared-memory FDs without sending the refreshed descriptors to shadow.
• A crash after such a remap could make shadow recovery use stale placeholder mappings with updated publisher/client-buffer metadata, risking corrupted recovery or crashes.
• Split-buffer option validation was asymmetric, allowing a later non-split publisher onto an already split-buffer channel, which could mix incompatible payload layouts.

Root cause

• Server::RemapChannel only updated local shared-memory FDs and did not replicate a refreshed channel create event to shadow.
• The remap path also failed to preserve the channel slot size for shadow metadata before any buffer allocation.
• ValidateOrSetSplitBufferOptions only rejected mismatches when the incoming option was true.

Fix

• Replicate remapped channel state to shadow with refreshed FDs and slot metadata.
• Preserve last_known_slot_size during remap.
• Require exact split-buffer option consistency once channel options are established.
• Added regression coverage for placeholder remap shadow state and mixed publisher split-buffer options.

Validation

• /tmp/subspace-tools/bin/bazelisk test //client:client_test --test_filter='ClientTest.RejectsMixedSplitBufferPublisherOptions'
• /tmp/subspace-tools/bin/bazelisk test //shadow:shadow_test --test_filter='ShadowRecoveryTest.ShadowTracksPlaceholderRemap'
• /tmp/subspace-tools/bin/bazelisk test //shadow:shadow_test //client:bridge_test