check SPV proof inner nodes not to be valid transactions

[SomberNight]

This attempts to implement protection against the attack recently described on the mailing list.

Before merging, I would like to measure the performance impact with a relatively large wallet on mainnet.
So far, I could only measure on testnet, but blocks are mostly empty there, so merkle proofs are short.

On testnet, I synced a wallet (A) that has ~350 txns.
Irrespective of this PR, the sync took ~10 seconds.
Without the patch, the verifier spent about 19 ms altogether in hash_merkle_root.
With the patch, the verifier spent about 66 ms altogether in hash_merkle_root.

On testnet, I synced a wallet (B) that has ~10.9k txns.
The sync took ~3.5 minutes. (with the patch, it was about 4 seconds longer, but that is within error probably; I only measured this once)
Without the patch, the verifier spent about 733 ms altogether in hash_merkle_root.
With the patch, the verifier spent about 2014 ms altogether in hash_merkle_root.

[SomberNight]

I've found this address on mainnet: 1NcXyC1m3XE1s3u9rBZawt3sCZkeGEgoH9,
which has made 9733 transactions, all in the last 6 weeks.

When syncing a wallet for the above address, I benchmarked the verifier.

Without the patch, the verifier spent about 647 ms altogether in hash_merkle_root.
With the patch, the verifier spent about 1946 ms altogether in hash_merkle_root.
_raise_if_valid_tx was called 109422 times.