We can call do_tb_phys_invalidate from an iocontext, which has
no per-thread tcg_ctx.  Move this to tb_ctx, which is global.
The actual update still takes place with a lock held, so only
an atomic set is required, not an atomic increment.

Resolves: https://gitlab.com/qemu-project/qemu/-/issues/457


Tested-by: Viktor Ashirov <vashirov@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>