tinydiffeq 2.3.0: port to nlls-gram 2.7.0 - #8

The GPU tests build state with a bare `jnp.asarray(1.0)`, which under the x64
conftest is float64 -- so the core ODE, SDE and DAE GPU paths had no
single-precision coverage at all. Four float32-end-to-end tests close that, at
a deliberately coarse bar: no crash, on device, finite, and correct to a few
digits. Tight float32 agreement is a CPU concern; the point here is that the
GPU kernels take the same paths and do not return nonsense.

One of them pins the dtype rule in dae.py, which reads z_dtype ALONE to pick
root_atol (1e-10 above 32 bits, else 1e-6): a float32 y with a float64 z gets
the tight bar, and both dtypes have to survive the round trip on device.

conftest also sets jax_default_matmul_precision=highest. XLA:GPU serves
float32 dot_general from TF32 tensor cores by default at a 10-bit mantissa,
and the float32 tests here check against closed forms with tolerances derived
from float32 eps -- the linear-exponential and Markov paths are matmul-heavy
enough that TF32 would fail them on GPU while passing on CPU. nlls-gram pins
the products it owns, but it cannot reach the matmuls in a caller's vector
field, which is exactly what these tests exercise. A no-op on CPU.

Co-Authored-By: Mecha Perla (Claude) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014og23CSBQfdHGNCfA21F8x

The two new float32 GPU tests read .xs[-1]. SaveAt defaults to t_1=True, so
xs is already the endpoint -- a scalar here, and IndexError on a GPU box. The
DAE cases passed only because their state is a 1-element array. Caught by the
3090 run, which is the point of having them.

Co-Authored-By: Mecha Perla (Claude) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014og23CSBQfdHGNCfA21F8x

Floors nlls-gram at the released 2.7.0 and drops the local editable source
override, so the published wheel resolves from PyPI instead of carrying a
path into a checkout.

2.2.0 and earlier are not compatible with nlls-gram 2.7.0: LMRootSolver
forwarded 26 constructor kwargs and 17 of them no longer exist, so a fresh
install of 2.2.0 against 2.7.0 fails at solver construction. 2.6.0 still
accepted them, which is why the break starts now.

LMRootSolver went from 26 fields to 6 in that port: the five this package
actually owns, which reach the nlls solve rather than its constructor, plus a
solver_options pass-through for the rare root that needs a non-default
algorithm. Verified against the published wheel: 251 passed, 15 skipped.

Co-Authored-By: Mecha Perla (Claude) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014og23CSBQfdHGNCfA21F8x