Concurrent map write from llb

[hinshun]

When marshalling more than one llb.State that share vertices.

fatal error: concurrent map writes
goroutine 181 [running]:
runtime.throw(0x23051fb, 0x15)
	/usr/local/go/src/runtime/panic.go:1112 +0x72 fp=0xc0005e4910 sp=0xc0005e48e0 pc=0x1035da2
runtime.mapassign_faststr(0x20d26e0, 0xc0003431d0, 0x22fccf8, 0x10, 0xc000452309)
	/usr/local/go/src/runtime/map_faststr.go:211 +0x3f7 fp=0xc0005e4978 sp=0xc0005e4910 pc=0x1015357
github.com/moby/buildkit/client/llb.addCap(...)
	/go/src/stash.corp.netflix.com/engtools/newt/vendor/github.com/moby/buildkit/client/llb/source.go:468
github.com/moby/buildkit/client/llb.(*ExecOp).Marshal(0xc0007a5080, 0x2597cc0, 0xc00019e008, 0xc000364540, 0xc00033f650, 0xc0005e58f0, 0xc0005e5948, 0xc0005e5918, 0xc000364540, 0xc00033f650, ...)
	/go/src/stash.corp.netflix.com/engtools/newt/vendor/github.com/moby/buildkit/client/llb/exec.go:235 +0xc72 fp=0xc0005e4ff0 sp=0xc0005e4978 pc=0x1ad4bf2

[sipsma]

This appears to be the same issue that was causing occasional panics in mergediff tests, i.e. as mentioned here.

The issue there appeared to be that the same llb.State was being used by distinct buildkit clients running in separate goroutines.