[qa] Ensure we don't generate a too-big block in p2sh sigops test #16464

There's a bug in the loop that is calculating the block size in the p2sh sigops test -- we start with the size of the block when it has no transactions, and then increment by the size of each transaction we add, without regard to the changing size of the encoding for the number of transactions in the block.

This might be fine if the block construction were deterministic, but the first transaction in the block has an ECDSA signature which can be variable length, so we see intermittent failures of this test when the initial transaction has a 70-byte signature and the block ends up being one byte too big.

Fix this by double-checking the block size after construction.

@jonasschnelli I believe this explains the test failures you've seen.

Looks good. utACK bf3be5297a746982cf8e83f45d342121e5665f80

For the record: I saw that feature_block.py failed (twice in the last days) on my custom CI while travis did not fail (https://bitcoinbuilds.org/index.php?job=2ff386f7-798c-4f9a-8233-dc62978cb175).

Thanks @sdaftuar for tracking it down.

tested ACK bf3be5297a746982cf8e83f45d342121e5665f80

Tested by applying this patch to master (fails) and this PR (passes):

-- a/test/functional/feature_block.py
+++ b/test/functional/feature_block.py
@@ -467,6 +467,8 @@ class FullBlockTest(BitcoinTestFramework):
         tx = self.create_tx(spend, 0, 1, p2sh_script)
         tx.vout.append(CTxOut(spend.vout[0].nValue - 1, CScript([OP_TRUE])))
         self.sign_tx(tx, spend)
+        while len(tx.vin[0].scriptSig) > 71:
+            self.sign_tx(tx, spend)
         tx.rehash()
         b39 = self.update_block(39, [tx])
         b39_outputs += 1

The patch grinds signatures until one of size <= 70 bytes is found, and the proceeds.

I've done some analysis of signature sizes produced by our python ECDSA implementation (#15826) and the OpenSSL code that we used before. Both use low-s by default and do not use RFC6979. In both cases, 70 and 71 byte sigs occurred just under 50% of the time and 69 byte sigs occurred ~0.4% of the time (sizes exclude the SIGHASH byte). That means that without this fix, we'd expect the test to fail around 1/250 - 1/200 times.