Replace libc printf with eyalroz/printf and add streaming status formatting. (#23694)

Integrates [eyalroz/printf](https://github.com/eyalroz/printf) as a
vendored dependency to replace all libc snprintf/vsnprintf calls across
the runtime. This gives us a consistent, portable printf implementation
across all platforms (including embedded and bare-metal targets) with no
platform-specific quirks.

Beyond the drop-in replacement, this extends the status formatting API
with streaming support:

- Add iree_status_format_to: callback-based streaming formatter that
avoids contiguous buffers and pre-measurement passes. The callback
returns bool for flow control, enabling early termination on buffer full
or allocation failure.

- Add iree_string_builder_append_status: single-pass status-to-builder
formatting via iree_status_format_to, resolving the
status/string_builder dependency cycle that previously required a local
copy in libhsa.c.

- Rewrite iree_status_fprint to stream directly to FILE* with zero heap
allocation (previously required a contiguous buffer).

- Rewrite iree_status_to_string as single-pass using string builder
instead of two-pass measure-then-format.

- Fix va_copy usage in status allocation: use C99 va_copy for the
two-pass measure-then-format pattern instead of calling va_start twice
on the same parameter (implementation-defined behavior).

- Convert Status::ToString (C++) and ApiStatusToString (Python bindings)
from two-pass iree_status_format to single-pass iree_status_format_to.

- Replace HIP dynamic_symbols.c allocate-format-free pattern with direct
iree_string_builder_append_status (eliminates intermediate heap
allocation).

- RISC-V toolchain fix: strip debug info from libgcc.a during CMake
configure. eyalroz/printf's long double support in %Lf/%Le compiles
`va_arg(args, long double)` which on RISC-V (128-bit quad precision)
generates references to `__extenddftf2`/`__trunctfdf2` soft-float
builtins. These are resolved from libgcc.a, but the GCC 12.2.0-compiled
objects in that archive contain DWARF debug info with relocation types
60/61 (in `.debug_rnglists` and `.debug_loclists` sections) that LLD
cannot parse. The pre-built RISC-V toolchain lacks compiler-rt builtins
as an alternative, so we strip the debug metadata from libgcc.a at
configure time using `llvm-objcopy --strip-debug`. This preserves all
code and data sections and is idempotent. The strip runs inside the
`RISCV_TOOLCHAIN_ROOT` guard so users with their own toolchains are
unaffected. Verified locally with the same toolchain used in CI
(`toolchain_iree_manylinux_2_28_20231012`).

Co-authored-by: Claude <noreply@anthropic.com>