[rootkiller6788]

tf.randomGamma(x, alpha, beta) was sampling with a mean of alpha * beta instead of alpha / beta.

The class stores the scale, i.e. 1/beta, in this.beta, but nextValue() finished with v = (1 / this.beta) * this.d * v - multiplying by the rate a second time. Every sample came out beta^2 too large. Since beta defaults to 1 this is invisible unless you pass a real rate, and the existing tests only assert the values are somewhere in [0, 30], which any scaling passes.

I noticed it while checking the sampler's mean against the distribution. Sample mean over 400k draws:

alpha=2,  beta=2   -> 3.9934  (should be 1.0)
alpha=2,  beta=0.5 -> 1.0007  (should be 4.0)
alpha=0.5,beta=3   -> 1.4971  (should be 0.1667)
alpha=2,  beta=1   -> 2.0035  (unchanged, default path)

So the fix is the single line, and it leaves the default beta=1 behaviour alone.

I also added the mean/std checks to rand_test.ts and random_gamma_test.ts, because a range check alone can't tell a correct scale factor from a wrong one. With the fix the seeded check gives mean 1.0020 / std 0.7114; against the old code the same check fails with 4.0079 vs 1.

I couldn't run the bazel test suite here (no bazel on this machine), so I verified by compiling the changed rand_util.ts on its own against a real seedrandom and comparing both versions, plus the table above.