also interested in knowing this, I have the same question/issue

AFAIK yes you can. And again AFAIK people have been simultaneously playing different synth implementations (ile OPL3+CrystalFM etc.) and mixing them together to get richer FM sounds.

Reads will cause a bus conflict but luckily all the parties involved drive same data on the bis and roughly at the same time so too there's no real harm happening. This is one reason why I have "write only" option for FM on my sound card design, useful in these situations or using as complement/replacement to exsiting FM chip.

Yes, there is a conflict on read, which is likely an AND of all cards signals. But is there actually a case/game where it is read from the FM-Synth and it breaks for configurations with multiple cards exposing the Synth at e.g. 0x388h ?

Probably not, only case where there can be trouble is if a clone chip is involved but one that has very significantly different timings and I doubt such exists.
AND function only happens with TTL type parts, but most of these devices are CMOS based or often use CMOS bus drivers and in this case there's no simple answer since both 0 and 1 are driven equally strong rather than 1 is driven much stronger than 0 as it is with TTL parts.

One concrete example of problems:
5-card build - resource considerations

I can't get my Aztech Sound Galaxy NX Pro (early SBPro 2.0 clone with real OPL3 - and Coxov/Disney Sound Source support) to coexist with a Mediavision Pro Audio Spectrum 16. Once the NXPro is in the system, the PAS16 simply doesn't do anything (even digital audio). The NXPro itself works flawlessly.

Problem seems to be that the PAS16 uses the AdLib I/O for a lot more than just FM, it's also what the whole card ID/detection is based upon. Oddly, an original SB16 present in the same system (CT2910 or CT2950) doesn't cause any problems, just that NX Pro.

So yes, it's real, but it's a bit of a corner case, and neither affects nor is caused by any original Creative Sound Blaster I tried.