Fix regression of the zygiskd refactor

[JingMatrix]

The condition for caching a clean mount namespace was modified by mistake.

In particular, without checking mounts whose root paths start with /adb/modules, NeoZygisk fails to clean the dex2oat mount of LSPosed.

To reproduce this issue reliably, one should try JingMatrix/Vector#429. I wasn't able to spot this regression with the master branch of JingMatrix/LSPosed, which was the cause of this regression.

[JingMatrix]

On my device, if LSPosed daemon starts in service.sh mode, then the mount namespace caching of NeoZygisk always happens before the LSPosed mounting of dex2oat. This is the core to explain why I failed to spot the regression, since LSPosed doesn't exist in the root mount namespace at the moment of caching.

Only starting LSPosed daemon in post-fs-data.sh mode, I can reproduce this regression in my device, as the dex2oat mounting now happens surely before the mount namespace caching of NeoZygisk.

Howover, on some other devices, even if LSPosed daemon starts in service.sh mode, its mounting of dex2oat still happens before the mount namespace caching. Hence, we must fix this regression for them.