Sure enough, SI buffer overrun is broken

I built the current sources with MAXNUMMESSAGES set to 32 in
src/include/storage/sinvaladt.h.  The regular regress tests
run OK, with just a few NOTICEs about 'cache state reset'
and 'SI buffer overflow' inserted in the normal outputs
(as you'd expect, if SI overrun occurs).

However, the parallel tests crash spectacularly, with weird errors
and Assert() coredumps.  Some of the unexpected messages in the
postmaster log are:

ERROR:  Relation 0 does not exist
NOTICE:  LockRelease: locktable lookup failed, no lock
TRAP: Failed Assertion("!(((file) > 0 && (file) < SizeVfdCache && VfdCache[file].fileName != ((void *)0))):", File:
"fd.c",Line: 817)
 

!(((file) > 0 && (file) < SizeVfdCache && VfdCache[file].fileName != ((void *)0))) (0)
NOTICE:  LockRelease: locktable lookup failed, no lock
TRAP: Failed Assertion("!(attnum <= 0 || (attnum - 1 <= tuple_type->natts - 1 && tuple_type->attrs[attnum - 1] !=
((void*)0) && variable->vartype == tuple_type->attrs[attnum - 1]->atttypid)):", File: "execQual.c", Line: 283)
 

!(attnum <= 0 || (attnum - 1 <= tuple_type->natts - 1 && tuple_type->attrs[attnum - 1] != ((void *)0) &&
variable->vartype== tuple_type->attrs[attnum - 1]->atttypid)) (0) [Not a typewriter]
 
TRAP: Failed Assertion("!(((file) > 0 && (file) < SizeVfdCache && VfdCache[file].fileName != ((void *)0))):", File:
"fd.c",Line: 817)
 

!(((file) > 0 && (file) < SizeVfdCache && VfdCache[file].fileName != ((void *)0))) (0) [Not a typewriter]

We have a problem.

I think Hiroshi was beating on this code recently --- Hiroshi,
do you recall anything you might have done that would affect
SI cache reset recovery?
        regards, tom lane