hs-v3: Close intro circuits when cleaning up the client descriptor cache

In legacy/trac#28970 (moved), one of the assert indicates that we are missing the descriptor object when the intro point circuit opened:

Jan 01 05:31:33.000 [warn] tor_bug_occurred_(): Bug: ../src/or/hs_client.c:624: setup_intro_circ_auth_key: Non-fatal assertion !(desc == NULL) failed. (on Tor 0.3.4.9 )
Jan 01 05:31:33.000 [warn] Bug: Non-fatal assertion !(desc == NULL) failed in setup_intro_circ_auth_key at ../src/or/hs_client.c:624. Stack trace: (on Tor 0.3.4.9 )
Jan 01 05:31:33.000 [warn] Bug:     /usr/bin/tor(log_backtrace+0x42) [0x56345cea26e2] (on Tor 0.3.4.9 )
Jan 01 05:31:33.000 [warn] Bug:     /usr/bin/tor(tor_bug_occurred_+0xb7) [0x56345cebd587] (on Tor 0.3.4.9 )
Jan 01 05:31:33.000 [warn] Bug:     /usr/bin/tor(hs_client_circuit_has_opened+0x2ca) [0x56345ce8027a] (on Tor 0.3.4.9 )

When a descriptor is removed from the client cache, the intro circuits aren't closed so there is a race where if it happens in the same main loop run that the client has an opened intro circuit for it, then it could lead to that assert.

Regardless of the cause of the assert or not, we should always close pending intro circuits when cleaning up a descriptor since once it opens, the client requires access to the descriptor object to complete the introduction (see setup_intro_circ_auth_key()).

Funny thought that we do that when we replace a descriptor from the client cache but not when we purge it...

This is a possible backport contender in order to avoid BUG() and failure of reachability client side.