[Fix] Keep Buffer input validation active under python -O

[morluto]

Part of #1.

Summary

Replace correctness-critical assertions at the public Buffer boundary with
typed exceptions that remain active under python -O.

caller tensors and configuration
                |
                v
     +------------------------+
     | public Buffer boundary |
     | type, shape, layout    |
     | CUDA and exact device  |
     +-----------+------------+
                 |
        valid ---+---> planning / CUDA launch
                 |
      invalid ---+---> typed Python exception

Problem

Python removes assert statements when optimization is enabled. Public checks
for tensor dtype, shape, contiguity, device placement, required arguments, and
zero-copy aliasing could therefore disappear before malformed inputs reached
CUDA-facing code.

Change

Add a shared public tensor validator covering:

• object type;
• dtype;
• shape;
• contiguity;
• CUDA placement;
• exact Buffer device.

Apply explicit validation to:

• dispatch hidden states, route weights, top-k indices, and per-expert counts;
• prefetch weight tensors;
• combine inputs and zero-copy aliases;
• full and intermediate gradient-reduction tensors;
• Buffer construction and context constraints;
• destroyed or uninitialized Buffer use.

The exception types distinguish the failure category:

• TypeError for incorrect Python types and tensor dtypes;
• ValueError for incompatible shapes, layouts, devices, arguments, and
  aliases;
• RuntimeError for destroyed or uninitialized buffers.

Scope

This PR covers the top-level Buffer API in api.py. Internal launch-time
assertions in planning, dispatch, combine, buffer allocation, and
gradient-reduction modules remain separate work under #1.

It does not include the separate num_sms policy or communication-plan
provenance changes.

Validation

• Focused validation under python -O — passed.
• Python compilation and formatting checks — passed.
• Added or updated invalid-input coverage for dispatch, combine, prefetch,
  gradient reduction, and zero-copy aliasing.