docs: make the README's adaptive and np.savez recipes work on every layer

[aksheyd]

Two README recipes failed as written on real models.

Adaptive. On layers with outliers, a block can need more than 8 bits, and adaptive.quantize raises ToleranceTooTightError. That's on purpose since #76, but the README didn't say so, so a loop over a model's layers stopped at the first outlier layer. Its np.std(weights) also failed on PyTorch tensors, since NumPy hands the call to Tensor.std, which doesn't take NumPy's arguments.

• the line now uses weights.std(), which works on NumPy arrays and PyTorch tensors, including ones that require grad or hold bf16. A plain list needs np.std
• after the list, a four-line retry with the error's smallest_tolerance, which always passes, since it's half an 8-bit step of the widest block. The README says it loosens every block, not just the one with the outlier

try:
    q = adaptive.quantize(weights, tolerance=0.1 * weights.std())
except ToleranceTooTightError as error:
    q = adaptive.quantize(weights, tolerance=error.smallest_tolerance)

On 1,000,000 Student-t weights with 3 degrees of freedom, a common stand-in for heavy tails, the first call raises and the retry works at 0.18 of a standard deviation.

np.savez. For an adaptive tensor, bits=q.bits saved None, which np.load refuses without allow_pickle=True. The recipe now saves every part and leaves out whichever of bits and block_bits is None, so one recipe works for every kind:

parts = dict(kind=q.kind, shape=q.shape, block=q.block, bits=q.bits, block_bits=q.block_bits,
             codes=q.codes, scales=q.scales, zero_points=q.zero_points, scale=q.scale.name)
np.savez("layer.npz", **{name: part for name, part in parts.items() if part is not None})
q = Quantized.from_parts(**np.load("layer.npz"))

Run verbatim through a real file, it round-trips all 12 combinations of kind and scale type, empty tensors included. The existing test now runs this recipe instead of branching on the kind.

#98 changed the next two bullets, so this branch has main merged in, keeping #98's wording. PyPI keeps each release's README as uploaded, so this is worth having in 0.3.0. Merges cleanly with #101. just lint, just test, and just python pass.