If I understand well, that would mean we should use the same seed for bitcoin and password, modified by a different passphrase to separate curves to not mix things. That would be better than my actual password policy.
No need for separate curve. Thanks to BIP32, the room of possible keys from single space is really huge. Of course passwords will use different branch than private keys for bitcoin, for this reason we introduced BIP43/BIP44.
I have read that if we leak a single bip32 private key, all other keys can be disclosed. (I don't understand exactly how but it seem to be possible)
This is true, but tranversing works only on the same level of HD tree branch.
However, Trezor never leaks private keys to computer. It only uses internal private keys to actually encrypt/decrypt values provided by computer. So this attack vector is not possible in Trezor.
This is anyway real attack vector for software which uses HD wallets, but offer importing/exporting private keys. There you must be sure you don't leak master public key AND single private key.
Again, this is not a valid case with Trezor.
I assume (but maybe i'm wrong) that if a compromised computer can have my encrypted password (blob in password manager), plus the decrypted password (sent by Trezor), it's possible to compute the private key
No, this is not possible.