feat: Compatibility with @lala/appraisal

[retraigo]

This PR adds two features:

• Better type for Shape.
• An overload for tensor() to accept a TensorLike as a parameter.

This provides smooth interop with @lala/appraisal's ML utilities (class to categorical, softmax to class, text vectorizers, etc).

Using appraisal with netsaur before this PR:

const y_pre = [1, 1, 3, 1, 2, 2, 2, 1, 1, 3, 3, 3, 2, 1];
const encoder = new CategoricalEncoder();
const y = encoder.fit(y_pre).transform(y_pre, "f32");

tensor(y.data, y.shape)

Usage after this PR:

const y_pre = [1, 1, 3, 1, 2, 2, 2, 1, 1, 3, 3, 3, 2, 1];
const encoder = new CategoricalEncoder();
const y = encoder.fit(y_pre).transform(y_pre, "f32");

tensor(y)

[load1n9]

LGTM