[OvOhao]

Fixes #130

What changed

Several functions in cgo/ffmpeg passed C.CString(...) directly as a call argument, so the C strings were never freed:

• HasEncoder, HasDecoder (ffmpeg.go)
• NewAudioEncoderByName, AudioEncoder.SetOption, AudioEncoder.GetOption (audio.go)
• the six av_opt_set_int calls in Resampler.Resample (audio.go)

Each string is now bound to a variable and released with defer C.free(...). The six av_opt_set_int calls go through a small avOptSetInt helper that does the same. <stdlib.h> is added to ffmpeg.h for C.free.

Why it is safe

None of the callees keep the pointer after returning:

• avcodec_find_encoder_by_name / avcodec_find_decoder_by_name only compare the name.
• avcodec_profile_name_to_int (in ffmpeg.h) only compares the name with strcasecmp.
• av_dict_get only reads the key; the returned entry points into the dictionary, not into the key.
• av_dict_set with flags == 0 copies both key and value with av_strdup.
• av_opt_set_int only uses the name to look up the option.

How it was tested

In a golang:1.22-bullseye container with Debian's FFmpeg 4.3 dev packages (libavformat-dev libavcodec-dev libavutil-dev libavresample-dev libswscale-dev), before and after this change:

• go build ./cgo/ffmpeg/ and go vet ./cgo/ffmpeg/: both pass.
• A small program that runs 20000 cycles of HasEncoder("aac"), HasDecoder("aac"), NewAudioEncoderByName("aac"), SetOption("profile", "LC"), SetOption("b", 128000), GetOption("b", &s) and Close(), measuring mallinfo().uordblks around the loop:

before: cycles=20000 malloc in-use delta=4483792 bytes (224.2 B/cycle)
after:  cycles=20000 malloc in-use delta=4144 bytes (0.2 B/cycle)

224 B/cycle matches 7 leaked strings per cycle at glibc's 32-byte minimum chunk. The Resampler path is covered by the build only.