feat(autodiff): add the generated tensor constant operator and reader

[AlekseiChirkovVention]

Introduce the single public representation of a generated tensor-valued constant operand, so
that a later expansion can multiply against a tensor of ones that actually exists in the graph.

Goal

Add FillOperator — a concrete zero-operand TensorOperator whose op_params is exactly
{fill, dtype, shape} — together with the FillDescriptor record and the shared
fill_descriptor reader, so every backend materializes a constant from one validated schema
instead of hand-parsing operator parameters.

A concrete operator type, rather than a parameter on an existing operator, is what lets
handler dispatch reject a backend that does not understand the construct before any handler
runs. Existing operators keep their operand contracts unchanged: no handler anywhere sees a
new operand count, a new parameter key, or a new shape convention.

Acceptance criteria

• FillOperator is concrete, declares route_name "fill", and adds no field beyond the
  base operator.
• fill_descriptor returns the declared fill, dtype, and shape from a node and from its
  OperationContext, as the same record.
• A well-formed fill node's output_typespec equals its descriptor's dtype and shape.
• A missing key, an extra key, a non-real or boolean fill, a non-integer or negative
  dimension, an absent or non-string dtype, a node carrying an operand, and a node whose
  operator is not a fill each raise a categorized error naming the node id.
• No bare KeyError, IndexError, TypeError, ValueError, or AssertionError escapes the
  reader for those inputs.
• The three names are reachable from tinychain.autodiff, present in __all__, and absent
  from the top-level tinychain package.
• "fill" is absent from the tracing capture allowlist and the error-category tuple is
  unchanged.

Test evidence

Level: unit. Tests first: the failing set landed in ce1a5e2, before any
implementation existed; 7a869c7 made it pass.

py/tests/test_autodiff_fill_contract.py — 17 test functions, 59 parametrized cases, all
passing. Coverage per criterion: the operator shape by three tests asserting route_name and
the declared field set; the reader by two tests comparing the node and context results; the
typespec agreement by one; the categorized-rejection table by two 15-case parametrizations
plus three targeted tests; the builtin-leak guarantee by a third 15-case parametrization; the
export surface by one test asserting __all__ membership, identity against the module, and
absence from the top-level package; and the two nothing-changed criteria by one test each.

Neighbouring autodiff suites: 182 passed. Broad non-integration suite: 973 passed, 9 failed,
1 skipped — all 9 failures pre-existing and unrelated to this branch (7 require the unbuilt
local backend, 2 are symbolic-style guards naming a file this branch does not touch).

Stack position

Stack position 1 of 8 — based on main.

This is the base of a stacked series; the seven branches that follow build on it in order, so
they cannot reach main before this one does.

Review notes

Known follow-ups, already recorded and deliberately not fixed here:

• The reader raises a bare TypeError for a hand-built OperationContext whose op_params
  is not iterable or whose operand list is not sized. Unreachable through the framework's own
  construction path, which always supplies a mapping proxy and a tuple. The guard is deferred
  because the same one must serve the expansion emitters that will share this module.
• The typespec-agreement criterion is currently asserted against the test's own helper. The
  real guarantee is the truthful-shape audit that lands with the expansion passes, where
  production code emits the typespec.