Are there any documented cases of HSM breaches anywhere or involving a CA?

Attackers had effective control over the DigiNotar CA before it was distrusted and eventually went bankrupt in, I think, 2011. They may not have been able to extract the keys from the HSM (this would probably require physical access) but they had the ability to cause issuance without accurate records kept so there's not a lot of practical difference.

Incidents at WoSign/ StartCom presumably involved malfeasance by key staff. I guess that doesn't count as a breach unless you'd call it a "Bank raid" if the manager just empties the vault into his own car and flees.

At Symantec they knew third parties had the independent ability to issue with any of their CAs but that was specifically contracted third parties (in particular a Korean firm named CrossCert) not just random people, it's just that issuance records weren't properly kept and oversight was inadequate. Again the ability to cause issuance isn't technically a breach, the keys stayed inside the HSM but it was possible to cause unrecorded issuance so there's not much moral difference.

tialaramex did a great job answering the second part of your question, so I'll take a swing at the first.

I don't know of any public cases where an org has disclosed that an external attacker exfiltrated key material from an HSM. That being said, there have been a number of disclosed vulnerabilities against HSMs/vendors that could allow this sort of attack to happen. CVE-2015-5464 is my favorite of these. There are also plenty of attacks that compromise the servers that talk to the HSMs, which usually would give an attacker the ability to perform arbitrary crypto operations using the keys in the HSM with no restrictions and little-to-no audit trail. I also know of attacks where the compromised "servers" are part of the HSM itself, but outside of the crypto/FIPS boundary.

IIRC, the Diginotar attack (used to make a fake certs and MITM *.google.com for many Iranians) involved replacing some of the dll files used to interface with HSMs. Dunno if it's confirmed that this is how the bad certs were made.