Adding tsit5 as a solver

[psv4]

Closes #194

Tsit5 is a Runge-Kutta method found to be more efficient than dopri5. In Diffrax and DifferentialEquations.jl, tsit5 is being recommended over dopri5 as a default non-stiff solver.

The coefficients for the dense output are functions of the interpolating point, unlike with all other methods present. I switched the mid --> _mid for all other RKAdaptiveStepsizeODESolver and added a default function mid() which returns the default coefficients. For tsit5, it takes in x which is the normalized intermediate point value and then returns the correct coefficients.

[psv4]

One test is failing regarding the seminorm but I am not sure how to address it.

======================================================================
FAIL: test_seminorm (norm_tests.TestNorms) (dtype=torch.float32, device='cpu', method='tsit5')
----------------------------------------------------------------------
Traceback (most recent call last):
  File "E:\Gitlab\torchdiffeq\torchdiffeq\tests\norm_tests.py", line 300, in test_seminorm
    self.assertLessEqual(seminorm_nfe, default_nfe)
AssertionError: 38 not less than or equal to 32

----------------------------------------------------------------------
Ran 22 tests in 332.048s

FAILED (failures=1)

[rtqichen]

Thanks for the PR. Some of the tests can be finicky, I will double check when I get some time.

I left a question about the self.mid change.