feat(nx): Implement lazy views and symbolic shapes

[tmattio]

This commit introduces two major architectural improvements to nx:

1. Lazy View Operations: View operations (reshape, permute, expand, etc.) are now lazy - they update metadata instead of copying data. Views are only materialized when operations need to access the underlying data.

2. Symbolic Shapes: Replace concrete int arrays with symbolic dimensions that can be bound at runtime. This enables shape-polymorphic kernels and dynamic batching.

Key changes:

• Add Symbolic_shape module supporting Static/Dynamic dimensions
• Add Lazy_view module tracking sequences of view transformations
• Update tensor type to use Lazy_view.t instead of View.t
• Refactor backend interface to accept symbolic shapes
• Update Native and Metal backends to handle lazy views
• Add ensure_materializable pattern for on-demand view materialization

Benefits:

• Memory efficiency: Avoid unnecessary data copies for view operations
• Performance: Enable view fusion and better memory access patterns
• Flexibility: Support dynamic shapes for batching and compilation
• Foundation: Prepare for advanced JIT optimizations in Rune

The changes maintain backward compatibility and NumPy semantics while bringing nx closer to modern tensor frameworks.

We once again draw significant inspiration from Tinygrad.