Handling reference block on transaction execution

After https://github.com/0xMiden/miden-client/pull/2100, the client only tracks the MMR peaks for the current sync height.

As mentioned [here](https://github.com/0xMiden/miden-client/pull/2100#discussion_r3157271113), I think there could be a problem with how we handle the `ref_block` on transaction execution.

`Client::execute_transaction` uses the current sync height as `ref_block` and passes it to the executor. The executor then builds the partial MMR through the data store. Since the client only stores peaks at the latest sync height, the MMR could not be consistent with the `ref_block`. If a concurrent `sync_state` advanced the chain between the caller capturing `ref_block` and the data store reading the peaks, the execution would fail.

I think there are two options here:

1. Assume a sync never runs during transaction execution. This would allow us to simplify the existing code by assuming the sync height is always the same during execution. It need to be documented (and ideally enforced).

2. Track historical peaks on the store. Allow `blockchain_checkpoint` to hold one row per chain tip we've seen, and look up peaks by `ref_block`. Since the client only accesses sync heigh, we only need to store the peaks for blocks that were added as chain tip, not for all intermediate blocks.

I think option 2 makes more sense and is probably the safest.

Look at the `Client::execute_transaction` method and how it uses the current sync height as `ref_block`. Examine the executor's MMR building logic via the data store. Review the `blockchain_checkpoint` storage to understand how peaks are tracked. The goal is to ensure MMR consistency with the `ref_block` during concurrent state syncs.

Written by the indexing model from the issue text.