[integer] Rewrite and fix type conversion from all integral scalar types

[forfudan]

Fix broken conversion from Int128 and add proper support for all integral types (Int128, Int256, UInt128, UInt256, platform-sized Int/UInt).

The old code cast everything through UInt64, silently truncating wide integers. The Int.MIN special-case was also fragile.

Now from_integral_scalar dispatches per dtype at comptime:

• Small unsigned types → single word
• UInt64 → 2-word split
• UInt128/UInt256 → loop with bit ops
• Signed ≤64-bit → unsigned subtraction for magnitude
• Int128/Int256 → division-based chunk extraction (right-shift on signed wide ints was broken)
• Platform Int/UInt → size_of dispatch

Removed standalone from_uint64 and from_uint128 (folded into from_integral_scalar). Updated sqrt call sites accordingly.

Added tests for Int.MIN/MAX edge cases, Int128/Int256 MIN/MAX, UInt128.MAX, UInt256, platform-sized types, and zero across all types.