define `Random.GLOBAL_RNG = TaskLocalRNG()` (#51388)
`GLOBAL_RNG` is a relic from pre-v1.3 when our global RNG was not
thread-safe:
```
const GLOBAL_RNG = MersenneTwister()
```
`GLOBAL_RNG` was never really specified, but was referred to in
docstrings (of `rand` etc.), and people had to use it in packages in
order to provide a default RNG to their functions. So we have to keep
defining `Random.GLOBAL_RNG` for the time being.
In v1.3, we didn't have anymore a unique global RNG, but instead one per
thread; in order to keep code using `GLOBAL_RNG` working, we got this
new definition as a clever hack:
```
struct _GLOBAL_RNG <: AbstractRNG end
const GLOBAL_RNG = _GLOBAL_RNG()
```
I.e. `GLOBAL_RNG` is just a "handle", which refers to the actual
threaded-RNG when passed to `rand` functions. This entails defining most
functions taking a `default_rng()` to also accept `_GLOBAL_RNG`. See
#41123, and #41235 which is not even merged.
But since v1.7, we have `Random.default_rng()` (our now official way to
refer to the default RNG) returning `TaskLocalRNG()`, which is itself a
"handle" (singleton). So we can as well redefine `GLOBAL_RNG` to be
`TaskLocalRNG()` instead of `_GLOBAL_RNG()`, with the advantage that all
the relevant methods are already defined for `TaskLocalRNG`.
The only expected difference in behavior is `seed!(GLOBAL_RNG, seed)`,
which previously would update `Random.GLOBAL_SEED` (a hack used by
`@testset`), but now is equivalent to `seed!(TaskLocalRNG(), seed)`,
which doesn't update this global seed. This precise behavior was never
documented (`GLOBAL_SEED` is purely internal), so this should be fine.
