[fix][ttl] Materialize tensor loop state before compute lowering (#527)

[brnorris03]

Problem description

Python kernels can express tensor recurrences by rebinding a tensor variable to an expression that reads its previous value:

acc = acc + x

The value after the loop must be the final recurrence value. Before this change, the compiler did not consistently represent and lower that tensor state through control flow, so compute lowering could not consume loops that produced tensor results.

Additive recurrences also have a performance requirement. They should preserve the L1 accumulation strategy that += already uses on a reserved block, instead of storing every intermediate accumulator value in a compiler-allocated DFB. After this PR, acc = acc + x and out_blk += x lower to the same in-loop accumulating ttl.store; += is now a special case of self-rebinding.

What's changed

• Added ttl-materialize-loop-state, a TTL transform that removes ranked-tensor scf.for iter args before compute lowering. Additive recurrences of the form ttl.add(%iter_arg, %contribution) (and the commuted form) lower to a pre-loop initial store and an in-loop accumulating ttl.store; non-additive recurrences lower through a compiler-allocated DFB state slot.
• Extended Python AST lowering to emit scf.for iter args for self-rebinding loop variables (including tuple targets like a, b = a + d, b + d) and resultful scf.if for tensor variables assigned in branches.
• Factored shared compiler-allocated DFB materialization helpers so ttl-insert-intermediate-dfbs and ttl-materialize-loop-state use the same allocation convention.
• Fixed TTKernelInsertL1Accumulation::precededByNonAccumulatingPack so a cb_reserve_back or cb_push_back for a CB already covered by a closer pack no longer aborts the scan. Loops that accumulate into multiple CBs now detect their pre-loop init packs and reconfig L1 acc as "enable" before the loop.
• Lit coverage for additive, commuted additive, unary and binary recurrences, mixed tensor states, scalar iter args, multiple final-result users, multi-use add fallback, conditional recurrence, and zero-trip loop semantics. Device coverage (dtype-parametrized bf16/f32) for additive, non-additive, tuple-target, three-accumulator, multi-tile-block, and zero-trip recurrences.
• Conditional tensor rebinding inside a loop is correct at the materialize-loop-state level but is left as xfail end-to-end: ttl-assign-dst does not yet descend into nested regions, tracked as Extend ttl-assign-dst to cover tile ops nested in scf.if / scf.for regions #587.

Fixes #527

[brnorris03]

Code just moved into this common file from TTLInsertIntermediateDFBs.cpp, no changes