analysis: keep annotated reports aligned with authoring syntax

Problem

The annotated output emitted by tilefoundry analyze --operands is not source-aligned with the authoring model. It prints lowered/expanded IR forms for operations that the tutorial author wrote in DSL syntax, so a reader cannot reliably map the report back to the model.

Reproduction

In the tutorial source, Stage4_WeightPrepared contains:

wq_local = tf.reshard(w_q, (1, HIDDEN, HIDDEN @ cta.head), "smem")

The corresponding annotated report prints an expanded form like:

v1 = reshard(w_q, layout=ShardLayout(
        layout=Layout((1, 256, 8, 32), None),
        attrs=(S(2),),
        mesh=cta,
    ), storage=smem)

Likewise, Stage5_CachePrepared authoring uses a Python slice:

k_cache[:, base : base + BLOCK, :, :]

but the report prints a generated IR operation:

v16 = slice(k_cache, (0, v15, 0, 0), sizes=(1, 128, 2, 32), strides=(1, 1, 1, 1))

Expected behavior

Provide one of these source-aligned forms in analysis output:

1. Preserve the original authoring expression and source location alongside each report record.
2. Add a stable source mapping from each generated IR call/value back to the authoring expression.
3. Clearly label the output as lowered IR and provide a source-level rendering for tutorial/agent use.

The key requirement is that reshard layouts and tensor slices must not look like syntax sugar or unrelated operations: the report should explain their relationship to the authoring model and preserve enough information for an agent to navigate back to the source.

Context

This was found while rendering docs/tutorial/authoring.ipynb into the installed tutorial Markdown. The generated report is numerically useful, but its current representation is misleading for a syntax/tutorial page.

Start with the tilefoundry analyze --operands entry point and the rendering path used for docs/tutorial/authoring.ipynb. Reproduce the report for Stage4_WeightPrepared and Stage5_CachePrepared, then trace how generated IR records relate to the authoring expressions. Done means the analysis output preserves or clearly maps reshard layouts and tensor slices back to source-level syntax and locations.