Fix 'math' tests to tolerate float imprecision (google#409)

The `math.tan(math.pi / 4) == 1` test and its negative pi counter part
fails under arm64 architecture.

Use `assert.true(near(...))` to tolerate float imprecision, like the
other 8 instances in the same test data file.

Tested on 24 combinations of OS x arch x Go version.
- {linux, darwin} x {arm64, amd64} x  go1.{16, 17, 18}

starlark/testdata/math.star

@@ -255,8 +255,8 @@ assert.fails(lambda: math.sin("0"), "got string, want float or int")
 # tan
 assert.eq(math.tan(0.0), 0)
 assert.eq(math.tan(0), 0)
-assert.eq(math.tan(math.pi / 4), 1)
-assert.eq(math.tan(-math.pi / 4), -1)
+assert.true(near(math.tan(math.pi / 4), 1, 0.00000000000001))
+assert.true(near(math.tan(-math.pi / 4), -1, 0.00000000000001))
 assert.eq(math.tan(inf), nan)
 assert.eq(math.tan(-inf), nan)
 assert.eq(math.tan(nan), nan)