[ShangkunLi]

Summary

This PR introduces an explicit type/value model for TileArray buffers and makes pattern-specific code responsible for replacement legality.

It adds:

• SynapseType and SynapseValue base classes;
• BufferType for shape and element type;
• Buffer for symbolic program arguments;
• BufferSlice for typed static selections such as A[:, k];
• Buffer-based TileArray, frontend lowering, compiler, library, and test APIs.

It also simplifies PatternReplacer: patterns now define their own applicability and semantic preconditions, while the shared replacement path adapts buffer types, stages the generated kernel, and verifies the resulting IR.

Example

A = synl.Buffer("A", synl.i32[3, 4])

column = A[:, 1]

assert column.source is A
assert column.type == synl.i32[3]

value = synl.load(column, tile=array[0, 1])

The slice remains a frontend value. It lowers directly to the original buffer operand and configured element offsets:

%value = "neura.load"(%A) {
  constants = array<i64: 1, 5, 9>
}